Download Proceedings of the conference - Slovenský národný korpus

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts

Computational linguistics wikipedia , lookup

Transcript
Slovenská akadémia vied
Jazykovedný ústav Ľudovíta Štúra
NLP, Corpus Linguistics,
Corpus Based Grammar Research
Fifth International Conference
Smolenice, Slovakia, 25–27 November 2009
Proceedings
Editors
Jana Levická
Radovan Garabík
2009
c by respective authors
The articles can be used under the
Creative Commons Attribution-ShareAlike 3.0 Unported License
Slovak National Corpus
Ľ. Štúr Institute of Linguistics
Slovak Academy of Sciences
Bratislava, Slovakia 2009
http://korpus.juls.savba.sk/∼slovko/
Table of Contents
Scientific Text Corpora as a Lexicographic Source
Larisa Belyaeva ...................................................................................................................... 19
The Corpus of Georgian Dialects
Marine Beridze and David Nadaraia ...................................................................................... 25
Corpus of Computational Linguistics Texts
Tatiana Bobkova, Mariia Kasianenko, Kuzma Lebedev, Valentyna Lukashevych, Pavlo
Petrenko, and Liubov Grydneva .............................................................................................. 35
“We Only Say We Are Certain When We Are Not”: A Corpus-Based Study of Epistemic
Stance
Vaclav Brezina ........................................................................................................................ 41
A Model for Corpus-Driven Exploration and Presentation of Multi-Word Expressions
Annelen Brunner and Kathrin Steyer ...................................................................................... 54
Text-Oriented Thesaurus Retrieval System for Linguistics
Natalia P. Darchuk and Viktor M. Sorokin ............................................................................. 65
From Electronic Corpora to Online Dictionaries (on the Example of Bulgarian Language
Resources)
Ludmila Dimitrova .................................................................................................................. 78
Evaluating Grid Infrastructure for Natural Language Processing
Radovan Garabík, Jan Jona Javoršek, and Tomaž Erjavec .................................................... 93
Synset Building Based on Online Resources
Ján Genči .............................................................................................................................. 106
Shallow Ontology Based on VerbaLex
Marek Grác ........................................................................................................................... 114
Multimodal Russian Corpus (MURCO): General Structure and User Interface
Elena Grishina....................................................................................................................... 119
Electronic Lexical Card Index for the Ukrainian Dialects (ELCIUD)
Pavlo Grytsenko, Olena Siruk, and Viktor M. Sorokin .......................................................... 132
Inflectional Entropy in Slovak
Adriana Hanulíková and Doug J. Davidson .......................................................................... 145
Exploring Derivational Relations in Czech with the Deriv Tool
Dana Hlaváčková, Klára Osolsobě, Karel Pala, and Pavel Šmerk ....................................... 152
On Epistemicity, Grammatical Person and Speaker Deixis in Polish (Based on the Polish
National Corpus)
Łukasz Jędrzejowski .............................................................................................................. 162
A Russian EFL Learner Corpus from Scratch
Olga Kamshilova .................................................................................................................. 167
Preliminary Analysis of a Slavic Parallel Corpus
Emmerich Kelih .................................................................................................................... 175
Operators for Extending and Developing an Utterance (Based on Operators of Concessive
Relation)
Jana Kesselová ..................................................................................................................... 184
Changes in Valency Structure of Verbs: Grammar vs. Lexicon
Václava Kettnerová and Markéta Lopatková ........................................................................ 198
Corpus-Based Analysis of Lexico-Grammatical Patterns (on the Corpus of Letters of
N. V. Gogol)
Maria Khokhlova and Victor Zakharov ................................................................................. 211
‘New/novelty’ Concept Set Dynamics as a Marker of Lexical and Grammatical Paradigm
Evolution for Psychology Sublanguage
Oksana S. Коzаk ................................................................................................................... 217
Methodological Foundations for Contrastive Model of Verb Valence
Ružena Kozmová ................................................................................................................... 222
Dictionary of Štúr’s Slovak
Ľubomír Kralčák ................................................................................................................... 235
Annotation Procedure in Building the Prague Czech-English Dependency Treebank
Marie Mikulová and Jan Štěpánek ........................................................................................ 241
Automatic Analysis of Terminology in the Russian Corpus on Corpus Linguistics
Olga Mitrofanova and Victor Zakharov ................................................................................ 249
Using Speech and Handwriting Recognition in Electronic School Worksheets
Marek Nagy .......................................................................................................................... 256
Composite Lexical Units as an Element of Lexicographical Historical Computer System
Irina Nekipelova ................................................................................................................... 266
IT: Moving Towards Real Multilingualism
Antoni Oliver and Cristina Borrell ....................................................................................... 279
Introduction of Non-Verbal Means of Communication in the Corpus of Live Speech
Tatyana Petrova and Olga Lys .............................................................................................. 287
MorphCon – A Software for Conversion of Czech Morphological Tagsets
Petr Pořízka and Markus Schäfer ......................................................................................... 292
Recent Developments in the National Corpus of Polish
Adam Przepiórkowski, Rafał L. Górski, Marek Łaziński, and Piotr Pęzik ............................. 302
Spoken Texts Representation in the Russian National Corpus: Spoken and Accentologic
Sub-Corpora
Svetlana Savchuk .................................................................................................................. 310
The Meaning of the Conditional Mood Within the Tectogrammatical Annotation of Prague
Dependency Treebank 2.0
Magda Ševčíková................................................................................................................... 321
The Creation of the Morphological Ambiguity Depository in Ukrainian
Olga Shypnivska and Sergij Starykov .................................................................................... 331
Frequency of Words and Their Forms in Contemporary Slovak Language Based on the
Slovak National Corpus
Mária Šimková and Miroslav Ľos ......................................................................................... 340
Analysis of the Means Expressing Strong ‘Necessity Not To’ in English and Czech Based on
General and Parallel Corpora
Renata Šimůnková ................................................................................................................. 349
Diatheses in the Czech Valency Lexicon PDT-Vallex
Zdeňka Urešová and Petr Pajas ............................................................................................ 358
A Corpus of Spoken Language and Its Usefulness in the Research on Language Contact
Marcin Zabawa ..................................................................................................................... 377
Vybudování databází na základě slovníku jako korpus
Miloud Taïfi and Patrice Pognan .......................................................................................... 389
Slovko (2001–2009)
Five Editions of the International Conference
Slovakia cannot boast of many linguistic events that have been organised on regular basis and focusing on one specific field. This kind of symposiums is actually
rather unique, in the present-day Slovak context only three of them are known
nationally: onomastic conferences are held in different parts of Slovakia, Banská
Bystrica hosts conferences on communication and annual Young Linguists’ Symposium covers all linguistic disciplines as well as interdisciplinary areas (in 2010 its
20th edition will be held). Moreover, this seminar once saw the early presentations of
some of the pioneers of Slovak computational and corpus linguistics (Emil Páleš)
and also hosted the first participants of Slovko 2001 (Karol Furdík, Jozef Ivanecký).
Although there has been only 5 events named Slovko, all of them of interdisciplinary nature dealing with areas of computational and corpus linguistics, the conference
gained the international character and its tradition seems to be well rooted.
2001 was the year of organization of the first Slovko conference, which was held
on October 26–27 (at that time still called Computer Processing of Slovak and
Czech and attended exclusively by Czech and Slovak lecturers and audience). This
symposium represented in the first place an event organised “with the aim to
improve mutual awareness and knowledge of people in Slovakia involved in the
issues of computers related to the language and vice versa: language related to computers” (A. Jarošová: Malá inventúra pred hľadaním spoločného jazyka. In:
Slovenčina a čeština v počítačovom spracovaní. Bratislava: Veda 2001, p. 7)1. Quoting the author and the main organiser on behalf of the Ľudovít Štúr Institute of
Linguistics, Slovak Academy of Sciences (the second organiser being Vladmír
Benko on behalf of the Faculty of Pedagogy, Comenius University), contemporary
Slovakia featured only fairly isolated islets of respective activities in different scientific disciplines and theoretically-applied contexts, out of which also efforts in the
area of artificial intelligence and cognitive linguistics revealed to be relevant.
However, neither results of the automatised processing of Slovak language data nor
relevant language data in terms of electronic corpus of texts in Slovak language
were available at that time. In comparison with the Czech Republic it was a diametrically different situation since computational linguistics there had been
intentionally developed as an autonomous field for more than 30 years and the 100
million token National Corpus of the Czech language has been made available on
the internet in 2000. Czech lecturers could therefore “offer the Slovak professional
public as well as students of linguistic and non-linguistic disciplines rather comprehensive overview of the work results in the area of Czech computer processing; this
language belongs in this respect among the European and in terms of different parameters also among world’s elite” (ibid). The seminar met also another of intended
aims: “carry on with the scientific, pedagogic and organisational work of Ján
Horecký who had been doing his best since the beginning of the 1960s to apply the
1
Available from http://korpus.juls.savba.sk/structure/dicts.sk.html
principles and methods of mathematical linguistics on the material of the Slovak
language” (ibid) and “who witnessed the revival of the computational linguistics in
the Institute of Linguistics when he initiated in 1988–1989 a project of data base of
the Slovak language, within which an idea arose to start to build a corpus” (ibid, p.
9). 14 papers altogether were presented, 5 of which by Czech authors, and consequently published in the proceedings. These proceedings were – at the same time
– partial result of the Ľ. Štúr Institute of Linguistics and Faculty of Pedagogy
involvement in the multinational project Trans-European Language Resources Infrastructure II – TELRI II, carried out as a coordinated action within the European
Commission programme INCO-COPERNICUS in 1999–2001.
The question whether the need of mutual informing of people involved in the
field of computational linguistics in Slovakia and the exchange of experience with
foreign specialists was only a one-time, temporal issue or whether it had a broader
context, was answered in 2003 when Vladimír Benko (as a chief organiser on behalf
of the Ľ. Štúr Institute of Linguistics and Faculty of Pedagogy) held an international
scientific seminar on 24–25 October, this time already called Slovko, subtitled as
Slavic Languages in Computer Processing. The event was attended by 54 people
interested in this area, 20 out of which were from Slovakia and the rest from abroad.
This edition of Slovko saw the first presentation of the Slovak National Corpus as an
actually existing electronic database of written texts covering the period 1955–2003
(in August 2003 it was made available for the public with its 26 million of tokens, in
December 2003 the second version of the Corpus was released, containing 166 million of tokens) as well as the team of researchers of new department of the Ľ. Štúr
Institute of Linguistics, created in mid-2002. Fledgeling corpus linguists lead by
Mária Šimková presented first of all the whole of Corpus Project and then partial
research solutions in segmentation, lemmatisation and morphological annotation of
texts of the Slovak language. The first Slovak morphological tagset presented at
Slovko was shortly afterwards publicly examined and reviewed by Slovak and foreign specialists and at the beginning of 2004 started a manual annotation of selected
texts of the Slovak National Corpus.
Due to the fact that no Slovak university or college offered either an autonomous
course of study or a specific seminar of computational and corpus linguistics, the
Slovak National Corpus department as a chief investigator of the State Research
Programme: Integrated Computational Processing of the Slovak Language for Linguistic Research Purposes organised regular lectures and seminars focused on these
two disciplines. A part of it was published in the proceedings entitled Insight into
the Slovak and Czech Corpus Linguistics. Ed. M. Šimková. Bratislava: Veda 2006.
208 p.2 As a natural conclusion of these and also of some other activities was the
organisation of Slovko on 10–12 November in 2005 and receiving the Slovak
Academy of Sciences Prize for Building Infrastructure in Science on 11 November
2005. The third international seminar Slovko had already a programme and organizing committee and thanks to aroused interest it was also extended to computer
treatment of Slavic and East European languages. It was attended by more than 60
interested researchers, which has represented the highest number of participants of
one edition of Slovko. Altogether 29 papers were presented in more-or-less homo2
Available from http://korpus.juls.savba.sk/publications/
genous sessions: spoken corpora, speech analysis and synthesis, computer lexicography, parallel and historical corpora, terminology, e-learning, which were later on
published in the proceedings Computer Treatment of Slavic and East European Languages (Ed. R. Garabík. Bratislava: Veda 2005. 246 p.).
Slovko in 2007 (explicitly held on October 25–27) was again organised by the
Slovak National Corpus team (namely Radovan Garabík, Jana Levická, and Mária
Šimková), this time as a solid fourth edition of a biannual international conference
focused on NLP and computational lexicography and terminology. More than 50
participants had an opportunity to attend 37 lectures and presentations; the proceedings entitled Computer Treatment of Slavic and East European Languages 2007 (Ed.
J. Levická, R. Garabík. Bratislava: Tribun 2007. 318 p.) were available upon arrival
at the conference. Beside the topics that have been fundamental for every Slovko
(corpus development including spoken corpora: data collection, annotation and processing), the foreground was reserved for theoretical issues of computational
lexicography and terminography, bilingual lexicography and terminography, cooccurrence analysis and pertinent collocations of lexicographical and terminographical relevance, new methods in data extraction and terminology mining from
corpora, terminology databases and terminology management systems and so forth.
Contributions of Slovak authors covering basic as well as applied research pointed
to revolutionary changes that Slovak computational and corpus linguistics in particular had seen in the recent period. For the sake of the example, we can mention
several results of the Slovak National Corpus team work: on January 1st 2007 began
the second phase of the realisation of the State Research and Development Programme: Building the Slovak National Corpus and Integrated Computational
Processing of the Slovak Language for the Linguistic Research Purposes, the 350
million token fully lemmatised and morphologically annotated general corpus was
made available for public. Beside other products (database of digitised lexicographic
and other linguistic works, parallel corpora), highest priority has been assigned to a
partial project of the Slovak National Corpus, the Slovak terminology database,
which was presented by its chief investigator Jana Levická, as well as the Slovak
Spoken Corpus whose technical features were shown by (in cooperation with Milan
Rusko) R. Garabík.
We decided to hold Slovko 2009 in the Congress Centre Smolenice of the Slovak
Academy of Sciences to celebrate sort of a jubilee edition since it is the fifth in a
row (all the previous editions took place in Bratislava but in different venues). Originally, the conference was to have been focused on parallel corpora, however, since
the Institute of the Czech National Corpus, Charles University, Prague, organised a
conference with the same topic on 17–19 September 2009 within their InterCorp
Project, we sought to come up with a different key issue that could be of immediate
interest in the Central-European context. Since the Institute of the Czech Language,
Academy of Sciences of the Czech Republic ceased to organise regular conferences
entitled Grammar & Corpora (hosted by František Štícha), the main conference
topic was in fact right at hand: corpus linguistics, more precisely corpus based linguistic research. In this respect this year’s Slovko has once again gathered more
Slovak and Czech participants, which can be viewed also as a reminder of the beginnings of Slovko. Presented contributions point to other topics and, at the same time,
give evidence of the progress in this field: majority of them do not deal with a
development of standard corpora of written or spoken texts but rather with the tools
of text analysis, building, and usage of language resources, linguistic components of
information systems, computer-aided translation systems, localisation and lexicography, computer-aided language learning etc. Similarly as the organisers of the first
Slovko participated in international projects, the Slovak National Corpus team and
Slovak computational and corpus linguistics have been involved in several international cooperation projects, especially in the 7th framework programme (EU FP7 INF
211983 Conceptual Modelling of Networking of Centres for High-Quality Research
in Slavic Lexicography and Their Digital Resources).
Five editions of Slovko reflect the corpus and computational linguistics in Slovakia, Czech Republic, as well as in neighbouring and more distant countries, namely
Austria, Belgium, Bulgaria, France, Croatia, Hungary, Germany, Norway, Netherlands, New Zealand, Poland, Russia, Slovenia, Spain, Ukraine have had their
participants so far. Approximately 130 participants altogether from 17 countries
have attended Slovko conference (however, some of them have participated
repeatedly, on regular basis). Overall, Slovko conferences hosted almost 300 participants that could listen to and/or read nearly 150 papers. We would like to express
our gratitude for excellent cooperation to all members of programme committees
and abstract peer reviewers. We also thank to all participants for creating fruitful
working and social atmosphere, to Czech colleagues in particular for they have been
showing us support from the beginning and attended regularly all Slovko’s editions.
We appreciate very much the work of all members of each organising committee
and hope that the flag will be handed on.
Mária Šimková
Jana Levická
Slovko (2001 – 2009)
Päť ročníkov medzinárodného podujatia
V rámci lingvistických podujatí na Slovensku nemáme veľa takých, ktoré by sa
uskutočňovali pravidelne a boli zamerané na jednu oblasť. Takéto podujatia sú naozaj
skôr výnimočné, v súčasnom slovenskom kontexte sú širšie známe asi tri: na rôznych
miestach Slovenska sa konávajú onomastické konferencie, v Banskej Bystrici
konferencie o komunikácii a celú lingvistiku, ale aj interdisciplinárne oblasti zahŕňa
každoročné Kolokvium mladých jazykovedcov (v r. 2010 sa už chystá 20. ročník), na
ktorom sa ako mladí kedysi prezentovali aj jedni z priekopníkov slovenskej počítačovej a korpusovej lingvistiky (E. Páleš) a prví účastníci Slovka 2001 (K. Furdík,
J. Ivanecký). Podujatí s názvom Slovko, ktoré sú zamerané interdisciplinárne na
oblasti počítačovej a korpusovej lingvistiky, bolo zatiaľ len päť, no ukazuje sa, že sa
založila dobrá tradícia s medzinárodným presahom.
Keď sa v r. 2001 pripravovalo a 26. – 27. októbra aj uskutočnilo prvé Slovko (ešte
pod názvom Slovenčina a čeština v počítačovom spracovaní a za účasti výlučne
českých a slovenských prednášateľov i poslucháčov), išlo predovšetkým o podujatie
organizované „so zámerom zlepšiť vzájomnú informovanosť ľudí zaoberajúcich sa na
Slovensku problematikou počítačov vo vzťahu k jazyku a naopak, jazyka vo vzťahu
k počítačom“ (A. Jarošová: Malá inventúra pred hľadaním spoločného jazyka. In:
Slovenčina a čeština v počítačovom spracovaní. Bratislava: Veda 2001, s. 7) 1. Ako
konštatovala autorka a hlavná organizátorka seminára za Jazykovedný ústav Ľ. Štúra
SAV (druhým organizátorom bol V. Benko za Pedagogickú fakultu Univerzity
Komenského), v tejto oblasti boli v tom čase na Slovensku značne izolované ostrovčeky aktivít prebiehajúcich v rôznych vedných odboroch a teoreticko-aplikačných
kontextoch, z ktorých sa ako relevantné ukazovali aj aktivity v oblasti umelej
inteligencie a kognitívnej lingvistiky. Neboli však k dispozícii výsledky automatizovaného spracovania slovenských jazykových dát ani relevantné jazykové dáta
v podobe elektronického korpusu textov slovenského jazyka. Išlo o diametrálne
odlišnú situáciu v porovnaní s Českou republikou, kde sa počítačová lingvistika
cieľavedome budovala ako samostatný odbor viac ako tridsať rokov a národný korpus
českého jazyka bol v r. 2000 zverejnený na internete v rozsahu 100 miliónov tokenov.
Českí prednášatelia preto mohli „poskytnúť slovenskej odbornej verejnosti a študentom z lingvistických aj nelingvistických odborov ucelenejší pohľad na výsledky práce
v oblasti počítačového spracovania češtiny, ktorá patrí v tomto smere medzi európsku
a vo viacerých parametroch aj medzi svetovú špičku“ (tamže). Seminár splnil aj ďalší
zo svojich cieľov: „nadviazať na vedeckú, pedagogickú a organizačnú prácu Jána
Horeckého, ktorý sa od začiatku 60. rokov 20. storočia usiloval uplatňovať princípy
a metódy matematickej lingvistiky na materiáli slovenského jazyka“ (tamže) a „ktorý
stál aj pri revitalizácii počítačovej lingvistiky v Jazykovednom ústave, keď v rokoch
1
Dostupný aj z http://korpus.juls.savba.sk/structure/dicts.sk.html
1988 – 1989 pripravil projekt bázy dát slovenského jazyka, v rámci ktorej sa začalo
uvažovať aj o budovaní korpusu“ (tamže, s. 9). Na seminári odznelo a v zborníku
z neho bolo publikovaných 14 príspevkov, z toho 5 od českých autorov. Publikovaný
zborník zároveň predstavoval jeden z výsledkov účasti JÚĽŠ SAV a PdF UK
v mnohonárodnom projekte Trans-European Language Resources Infrastructure II –
TELRI II, ktorý sa ako súbor coordinated action uskutočnil v rámci programu
Európskej komisie INCO-COPERNICUS v r. 1999 – 2001.
Otázka, či potreba vzájomnej informovanosti pracovníkov v oblasti počítačovej
lingvistiky na Slovensku a výmena skúseností s odborníkmi zo zahraničia bola iba
jednorazová, dočasná, alebo má širší kontext, bola zodpovedaná v r. 2003, keď
V. Benko (ako hlavný organizátor za JÚĽŠ SAV i PdF UK) uskutočnil 24. – 25.
októbra medzinárodný odborný seminár už pod názvom Slovko, v podtitule
špecifikovaný na Slovanské jazyky v počítačovom spracovaní. Podujatia sa zúčastnilo
54 záujemcov o túto oblasť, z toho 20 zo Slovenska a 34 zo zahraničia. Na tomto
Slovku sa prvýkrát prezentoval Slovenský národný korpus – ako už reálne existujúca
elektronická databáza písaných textov z r. 1955 – 2003 (od augusta 2003 verejne
prístupná v rozsahu 26 miliónov tokenov, v decembri 2003 bola zverejnená druhá
verzia v rozsahu 166 miliónov tokenov) i ako kolektív pracovníkov nového oddelenia
JÚĽŠ SAV, založeného v podstate v polovici r. 2002. Začínajúci korpusoví lingvisti
pod vedením M. Šimkovej prezentovali jednak celý korpusový projekt, jednak
čiastkové riešenia pri segmentácii, lematizácii a morfologickej anotácii textov
slovenského jazyka. Prvá podoba slovenského morfologického tagsetu predložená na
Slovku bola zakrátko osobitne verejne oponovaná domácimi aj zahraničnými
odborníkmi a začiatkom r. 2004 sa rozbehla ručná morfologická anotácia vybraných
textov SNK.
Vzhľadom na to, že na žiadnej vysokej škole na Slovensku neexistoval samostatný
odbor a ani sa nevyučoval samostatný predmet počítačovej a korpusovej lingvistiky,
oddelenie SNK JÚĽŠ SAV ako riešiteľ štátnej úlohy Integrated Computational
Processing of the Slovak Language for Linguistic Research Purposes organizovalo
pravidelné prednášky a semináre z tejto oblasti. Časť z nich bola publikovaná
v zborníku Insight into the Slovak and Czech Corpus Linguistics. Ed. M. Šimková.
Bratislava: Veda 2006. 208 s.2 Prirodzeným vyústením týchto a ďalších aktivít bolo
zorganizovanie Slovka 10. – 12. novembra 2005 a prevzatie Ceny SAV za budovanie
infraštruktúry pre vedu 11. novembra 2005. Tretí medzinárodný seminár Slovko už
bol obohatený o vedecký výbor a organizačný výbor a vzhľadom na prejavený záujem
bol rozšírený na Slovanské a východoeurópske jazyky v počítačovom spracovaní.
Zúčastnilo sa ho vyše 60 záujemcov, čo bol a stále je najvyšší počet účastníkov
jedného Slovka. Celkovo odznelo 29 príspevkov vo viac-menej homogénnych
blokoch: hovorené korpusy, analýza a syntéza reči, počítačová lexikografia, paralelné
korpusy, historické korpusy, terminológia, e-learning a pod. Tieto vyšli v zborníku
Computer Treatment of Slavic and East European Languages (Ed. R. Garabík.
Bratislava: Veda 2005. 246 s.).
2
Dostupný aj z http://korpus.juls.savba.sk/publications/
Slovko v r. 2007 (25. – 27. októbra) organizoval kolektív SNK (najmä R. Garabík,
J. Levická, M. Šimková) už ako stabilizovanú 4. bienálnu medzinárodnú konferenciu
so zameraním na NLP a počítačovú lexikografiu a terminológiu. Vyše 50 účastníkov
si malo možnosť vypočuť 37 príspevkov a prezentácií, zborník Computer Treatment of
Slavic and East European Languages (Ed. J. Levická, R. Garabík. Bratislava: Tribun
2007. 318 s.) mali všetci k dispozícii priamo na podujatí. Okrem tém, ktoré sú ako
základné súčasťou každého Slovka (tvorba korpusov vrátane hovorených: zber dát,
anotácia a spracovanie), sa do popredia dostali teoretické otázky komputačnej
lexikografie a terminografie, bilingválna lexikografia a terminografia, kookurenčná
analýza a lexikograficky alebo terminograficky relevantné kolokácie, nové metódy
v extrahovaní dát a získavanie terminológie z korpusov, terminologické databázy
a systémy terminologického manažmentu a pod. Príspevky slovenských autorov,
pokrývajúce základný aj aplikovaný výskum, poukázali na prevratné zmeny, ktoré
slovenská počítačová a najmä korpusová lingvistika zaznamenala za posledné
obdobie. Napr. z produkcie SNK, ktoré 1. januára 2007 začalo 2. etapu riešenia štátnej
úlohy Budovanie Slovenského národného korpusu a elektronizácia jazykovedného
výskumu na Slovensku, bol verejnosti k dispozícii hlavný korpus v rozsahu 350
miliónov tokenov s plnou lematizáciou a morfologickou anotáciou na báze vlastného
tagsetu. Okrem ďalších produktov (databáza elektronických lexikografických a iných
lingvistických diel, paralelné korpusy) sa prioritným čiastkovým projektom SNK stala
Slovenská terminologická databáza, ktorú na konferencii predstavila jej hlavná
riešiteľka J. Levická, a budovať sa začal aj Slovenský hovorený korpus, ktorého
technické parametre prezentoval (v spolupráci s M. Ruskom) R. Garabík.
Slovko 2009 sme ako polojubilejné, piate v poradí umiestnili do Kongresového
centra SAV v Smoleniciach (všetky predchádzajúce boli v Bratislave, a to na
viacerých miestach). Hlavným zameraním mali byť pôvodne paralelné korpusy, ale
keďže Ústav Českého národného korpusu FF UK v Prahe organizoval takto profilovanú konferenciu v súlade s cieľmi svojho projektu InterCorp 17. – 19. septembra
2009, hľadali sme inú hlavnú tému, ktorá by mohla byť aktuálnou v stredoeurópskom
priestore. A keďže Ústav pro jazyk český AV ČR prestal organizovať pravidelné
konferencie Grammar & Corpora (pod vedením F. Štíchu), priam sa nám ponúkla ako
hlavná téma korpusová lingvistika, resp. korpusovo zamerané lingvistické výskumy.
V tejto súvislosti je na tohtoročnom Slovku opäť viac slovenských a českých
účastníkov, čo je aj istá pripomienka východiskovej situácie Slovka 2001. Posun
v tejto oblasti nielen v našich krajinách naznačujú aj ďalšie prezentované témy: už to
nie sú otázky budovania štandardného korpusu písaných či hovorených textov, ale
viac nástroje textovej analýzy, tvorba a využitie jazykových zdrojov, lingvistické
zložky informačných systémov, preklad s počítačovou podporou, lokalizácia a lexikografia, didaktika vyučovania cudzích jazykov s počítačovou podporou a pod. A tak
ako organizátori prvého Slovka boli členmi medzinárodných projektov, aj Slovenský
národný korpus a slovenská počítačová a korpusová lingvistika sa v súčasnosti
prezentuje vo viacerých medzinárodných spoluprácach, predovšetkým v 7. RP (EU
FP7 INF 211983 s nazvom Conceptual Modelling of Networking of Centres for HighQuality Research in Slavic Lexicography and Their Digital Resources).
Päť ročníkov Slovka je odrazom vývinu počítačovej a korpusovej lingvistiky na
Slovensku, v Čechách, v okolitých i vzdialenejších krajinách, z ktorých sú doteraz
zastúpené Belgicko, Bulharsko, Francúzsko, Gruzínsko, Holandsko, Chorvátsko,
Maďarsko, Nemecko, Nórsko, Nový Zéland, Poľsko, Rakúsko, Rusko, Slovinsko,
Španielsko, Ukrajina. Zahraničných účastníkov bolo na všetkých konferenciách
Slovko spolu asi 130 (nie sú to unikátne výskyty, viacerí sú našimi pravidelnými
hosťami) zo 17 krajín. Spolu so slovenskými účastníkmi sme privítali na týchto podujatiach takmer 300 záujemcov, ktorí si mali možnosť vypočuť a/alebo prečítať takmer
150 príspevkov. Za vynikajúcu spoluprácu ďakujeme všetkým členom vedeckého
výboru a posudzovateľom abstraktov. Ďakujeme všetkým účastníkom za vytváranie
tvorivej pracovnej i spoločenskej atmosféry, osobitne českým kolegom, ktorí nás od
začiatku podporovali a pravidelne sa zúčastňovali aj Slovka. Ďakujeme všetkým
členom doterajších organizačných výborov a dúfame, že štafeta pôjde ďalej.
Mária Šimková
Jana Levická
Scientific Text Corpora as a Lexicographic Source
Larisa Belyaeva
Herzen State Pedagogical University of Russia
Saint-Petersburg, Russia
Abstract. Nowadays applied lexicography is one of the special branches of
applied linguistics, its task is creation and updating different automated and
automatic dictionaries and databases, which are problem and subject oriented.
Completeness and adequacy of lexicographic systems to a considerable extent
determine the level and reliability of information and knowledge extraction
from the texts of various composition, structure and assignment.
Modern approach to dictionary creation assumes preliminary formation and use
of parallel corpora of modern texts, which can be considered as a database for
solving not only research tasks, but practical lexicographic tasks as well. If we
use a full-text parallel corpora as a lexicographic base it is necessary to expand
them with a corpora of machine translation results, analysis and comparison of
these corpora will make it possible to allocate such lexical units, which should
be considered as dictionary entries. The main problem is to establish the boundaries and structures of these lexical units.
1 Applied lexicography and scientific texts
Nowadays applied lexicography is one of the special branches of applied linguistics,
its task is creation and updating different automated and automatic dictionaries and
databases, which are problem and subject oriented. Completeness and adequacy of
lexicographic systems to a considerable extent determine the level and reliability of
information and knowledge extraction from the texts of various composition, structure
and assignment.
Modern approach to dictionary creation assumes preliminary formation and use of
parallel corpora of modern texts, which can be considered as a database for solving
not only research tasks, but practical lexicographic tasks as well. Written text corpora,
as a rule, include the texts as they are, as well as text layouts: format boundaries and
features, parsing results necessary for establishing morphological characteristics of
lexical units. These texts can be used serve for concordance creation, word and collocation lists in case of monolingual corpora, as well as for creation multilingual
lexicons and concordance if we have parallel corpora.
If we use full-text parallel corpora as a lexicographic base it is necessary to
expand them with a corpora of machine translation results assumes, analysis and comparison of these corpora will make it possible to allocate such lexical units, which
should be considered as dictionary entries. The main problem is to establish the
boundaries and structures of these lexical units.
20
Larisa Belyaeva
Noun phrases are objects of special research in both theoretical and applied
aspects. Such phrases are functionally equivalent to a word, but at the same time they
represent a convolution of a sentence, i.e. they are rather units of syntax, not lexicon.
Thus we can assume that internal structure of a noun phrase correlate with internal
dependencies structure of the appropriate sentence. The problem is to find a procedure
or approach to recognize this structure in a convolution.
One of the most serious problems of English scientific text analysis and machine
or human translation is determination of dependency structure in noun phrases (NP).
The problem is related to the fact than when translating from English to any inflectional language we should know the relation structure between the NP components.
In scientific text simple noun phrases are multicomponent units with large number
of attributive elements in preposition to the NP head. Being dependent members of a
sentence, these phrases form one syntactic group with its head, syntactic function of
which coincides with syntactic function of the phrase as a whole. The report investigates the possibility of using corpora information for solving the problem of NP
structure recognition and translation.
Written text corpora, as a rule, include the texts as they are, as well as text layouts:
format boundaries and features, parsing results necessary for establishing morphological characteristics of lexical units. These texts can be used for concordance
creation, word and collocation lists in case of monolingual corpora, as well as for creation multilingual lexicons and concordance if we have parallel corpora.
2 Noun groups in a scientific text
During the translation process the text analysis is based on formal parsing and
semantic analysis. Both of these processes are based on our possibility to understand
the surface structure of a sentence and semantic relations between its components.
Noun phrases (NP) are the objects of special research in both theoretical and
applied aspects. Such phrases are functionally equivalent to a word, but at the same
time they represent a convolution of a sentence, i.e. they are rather units of syntax, not
lexicon. So we can assume that internal structure of a noun phrase correlate with
internal dependencies structure of the sentence. The problem is to find a procedure to
recognize this structure in a concise form of a NP.
Thus one of the most serious problems of English-Russian scientific text analysis
and machine or human translation is determination of dependency structure in NPs.
The problem is related to the fact than when translating from English to any inflectional language we should know the relation structure between the NP components. In
scientific text simple (without preposition) NPs are multi-component units with large
number of attributive elements in preposition to the NP head element. Being dependent members of a sentence, these phrases form one syntactic group with its head
element, syntactic function of which coincides with syntactic function of the phrase as
a whole.
Scientific Text Corpora as a Lexicographic Source
21
In machine translation procedure the structure of each NP and its boundaries are to
be determined at the sentence analysis step, thus the task of NP translation is to be
performed in the framework of the following operations:
• Establishing the head element of the English noun phrase;
• Establishing semantic and syntactic structure of the English NP;
• Finding semantic, syntactic and lexical structure of the Russian NP;
Since a NP is a sentence convolution, a compression of this structure, such
external simplification of both the structure and the form causes the NP semantic
complication. The markers of relations between actual components and types of relations between elements, which sentence shows with the help of different means, are
absent in the English NP.
Basic noun phrases in English are two-element combinations with a head noun,
frequency of which in scientific text exceeds frequency of three-element combinations in three times. (see Tables 1-3).
Rang
1
2
3
4
5
6
7
8
9
10
NP type
T+N
A+N
N1 + N2
A + N1 + N2
H+N
M/S + N
E+N
N1/A + N2
S+N
N1 + N2 + N3
NP frequency
2253
1584
1368
485
241
215
208
199
195
173
Accumulated frequency
2253
3837
5205
5690
5931
6146
6354
6553
6748
6921
%
23,07
39,29
54.32
58.26
60,74
62„93
65,06
67,09
69,09
70.87
Table 1. Fragment of frequency list of English NP structures
No
1
2
3
4
5
6
7
Total
NP length
2
3
4
5
6
7
8
Number of different models
1516
674
207
51
20
2
5
2475
NP frequency
3457
1053
380
61
164
2
5
5122
Table 2. Frequency of English noun phrase length in technical texts
(subject domain “Seismic protection”)
22
Model
number
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
Larisa Belyaeva
Model
Length
Frequency
Number of different NP
A+N
A+PII+N
N1+N2
N1+N2+N3
A/N1+N2
N1+G/N2
A1+A2/N1+N2
A+G/N2 +N2
A1+A2+N
A+N1+N2
PII+A+N
PII+N
A1+N1+A2/N2+N3
A1+C+A2+N
A1+A2/N1+N2+N3
2
3
2
3
2
2
3
3
3
3
3
2
4
4
4
1474
9
1407
248
71
24
10
7
151
292
25
170
3
15
6
748
6
530
128
47
18
10
4
104
172
20
73
2
9
7
Table 3. Models of noun phrases in technical texts
However external simplicity of the most frequent English NP structures is misleading. The fact is that this simplicity could be the result of initial noun phrase or
sentence compression. Such compression, formal simplification of NP structure leads
to its semantic complication.
Pursuant to these, formation of noun phrases in a real text is based on either merging noun phrases and separate lexical units in a new, more complicated nominative
construction, or on condensing multi-component NPs at the expense of deletion of the
units which are implicitly obvious.
Formation of a multi-component NP in a text is realized in any of two ways
depending on the type of nomination: either as a process of step-by-step complication
and specification of the object nomination (gradual complication of a noun phrase
with addition of its head element characteristics), or as a process of sequential convolution. This process is realized successively on several levels:
Level 1: transfer from a complex noun phrase to a simple one due to element
inversion.
Level 2: elimination of component duplication in a new NP.
Level 3: coordination of semes and elimination of components with duplicated
semes.
Referential status of noun phrases in scientific text permits us assume that author’s
attitude on information transfer and its understanding requires explication of relations
within the text. Analysis of texts in different subject domains had shown, that occurrence in the text a NP with length more than 2 elements is followed by occurrence a
Scientific Text Corpora as a Lexicographic Source
23
2-compound NP in the nearest context, within the limits of 2-3 sentences or combination of title, key words and abstract. Hence, at human translation we can use this
situation as a key for structure diagnostics. At MT we need to create a special text
translation memory.
The peculiarities of NP formation in the text are to be analyzed as in the following
example:
Connection of two two-element NPs into a new one which results in occurrence of
• four-component NP, the structure of which depends on the structure of merging
NPs, for example, if two groups of adjective + noun type merger the NP which
plays the role of an attribute is embedded in the position of the head element of the
first NP attribute:
indirect method + seismic analysis  indirect seismic analysis method
adult learner + second language  adult second language learner
• three-component NP in case, when one of the elements in two initial NP coincides,
for example
mental processing + processing operation 
mental processing operation
with establishing direct relations between (in this case) adjective mental and noun
processing,
• three-component NP in case, when semantics of one of the NP elements is supported as a part of a new noun phrase by the semes of other NP components, for
example, merger of NPs
communicative method + language learning
results in occurrence of a new noun phrase
communicative language learning
• three-component NP in case, when semantics of one of the elements is implied as
a part of a new noun phrase at the expense of extralinguistic information of the
domain, for example, merger of NPs
seismic stability + direct analysis
results in formation of the NP
seismic stability direct analysis,
which in the text may be convoluted up to three-component NP
seismic direct analysis
The cases of noun phrase transformation considered here under the condition of text
coherence and cohesion do not show all possible variants of their development in a text,
however, they give the basis for consideration of possible translation of a noun phrase with
high degree of structure compression. Besides the research conducted permits to show,
that exactly two-element noun phrases present special difficulties at their analysis and
translation.
To solve the problem of such NP translation we can see only two approaches which
can be used both in human and machine translation MT).
24
Larisa Belyaeva
The first approach includes modeling the knowledge base of the domain in question
(in the framework of a MT system) or appealing to such factual knowledge of the translator. In case of MT this approach is based on vast investigations of the possible relations
between both the main concepts of the domain and the items of the linguistic data base.
Creation of such a thesaurus or a semantic net is not only extremely laborious but spaceconsuming. But the most serious disadvantage of this approach is that an unambiguous
solution of the problem sometimes can't be achieved. For example, for a noun phrase constant amplitude deformation cycle a semantic network would show relations between the
nodes constant and amplitude, constant and deformation, constant and cycle and it is
impossible to use this information to establish the dependencies structure of the NP both in
human and machine translation.
The second approach could be more formal: we can use the information, which
can be received on the basis of the whole text analysis. This approach seems more
expedient as it is based on the formal indications of the author's intentions which are
reflected both in the text structure and in the composition of different NP with the
same constituents.
Investigations of text structure in terms of NP composition in different subject
domains (medicine, seismic isolation, space systems, power plants construction, language teaching etc) had shown that dependency structure of NP with 3 or more
constituents can be obtained from the nearest context: a 2-component NPs would
show the accurate relations relevant for this special text.
This means that for English-Russian MT we need a special means for noun phrase
analysis and transfer within the text boundaries, something like Text Translation
Memory, which could store the history of NP development.
Noun phrases are objects of special research in both theoretical and applied
aspects. Such phrases are functionally equivalent to a word, but at the same time they
represent a convolution of a sentence, i.e. they are rather units of syntax, not lexicon.
Thus we can assume that internal structure of a noun phrase correlate with internal
dependencies structure of the appropriate sentence. The problem is to find a procedure
or approach to recognize this structure in a convolution.
One of the most serious problems of English scientific text analysis and machine or
human translation is determination of dependency structure in noun phrases (NP). The
problem is related to the fact than when translating from English to any inflectional language we should know the relation structure between the NP components.
In scientific text simple noun phrases are multicomponent units with large number
of attributive elements in preposition to the NP head. Being dependent members of a
sentence, these phrases form one syntactic group with its head, syntactic function of
which coincides with syntactic function of the phrase as a whole. The report investigates the possibility of using corpora information for solving the problem of NP
structure recognition and translation.
The Corpus of Georgian Dialects
Marine Beridze and David Nadaraia
Arn. Chikobava Institute of Linguistics
Georgian Academy of Sciences, Tbilisi, Georgia
Abstract. The Georgian Language world is represented by three Kartvelian
languages and more than 25 dialects. The project “The Linguistic Portrait of
Georgia” is aimed to associate the problem of documentation and researching of
the Georgian dialects to the achievements of the corpus linguistics. The corpus
is now under development, in which quite vast textual collection is integrated. It
involves all the dialectal texts published during the last 100 years, archive
material obtained in all the dialectological field expeditions which took place in
the second half of the last century, and additionally dialectal texts recorded by
the project group in Georgia and its neighboring countries (Azerbaijan, Iran).
1 Introduction
The idea of corpus presentation and learning of the dialectal data by means of modern
technologies occurred in early researches based on the principles of corpus linguistics,
and seemed quite natural. Dialectology in its nature has always been more “corpusal”
among the other branches of Linguistics (as based on the specially gathered textual
material) and with its methods of description and investigation it has always
“required” more implementation technological progress achievements – Transition
from dialectographic paper to the electro-digital carriers was the result of following
the technical progress step by step. At the very beginning of corpus Linguistics
development, most of the methods of the new discipline (e.g. preparation of texts
collection, experience in textual and meta-textual annotation and so on) were familiar
for dialectology. Hence, relation between dialectal data and using the computer
system became available from the 70-ies of the last century (Gordon R. Wood 1969).
When “the great corpus construction works” – creation of national languages
corpora – began, this became important trigger to develop corpus dialectology.
According to the aim of the creators, several conceptions of processing of the
dialectal material data shaped out. Some researchers represent the dialectal data as
“deviations” from the literary language; that is why they describe only that part of
the data which differs from the “standard” language; others consider a dialect as a
model of the whole cultural-communicational area and treat each component more
carefully (Kryuckova, Goldyn 2008); quite different is the approach of those
investigators and creators, who create the corpora in the format of “sustaining the
languages under the risk of disappearing”, aimed to documental investigation of
small data of such languages, and so on.
26
Marine Beridze and David Nadaraia
The corpus of the Georgian dialects represents an attempt to create the most likely
model of the diverse language portrait of Georgia. The modern literary Georgian is
based on Kartvelian languages and their dialects; thus, it is impossible to investigate
the Georgian Language without considering data of the dialects. It should be noted
that in addition to the scientific importance (linguistic and culturological studies),
such corpora seem one of the best means to prevent the threats to the existence of
some languages and language products. The new millennium set this problem in new
light. The powerful wave of globalization rapidly changes the conceptions having
been established during centuries, the concept of the language security among them.
The State status, media, the education system and other social components can no
more provide the capable means of the language security “instituted” in the XX
century as guarantors. Today, any language, not having proper reflection of its
characteristics stored in computer system, or strong links with the global language
space or say it otherwise – do not have a translation system, strong textual massive
and tools of corpus investigation based on this, or do not have hyper textual thesauri
for synchronic and diachronic analysis of the lexis fund and so on, any such language
is under great risk of disappearing.
Unfortunately, in those years when the other countries actively strove to meet one
of the main challenges of the epoch – to document and investigate language events
documentation and investigation with modern technologies i.e. to develop corpus
linguistics, our country was dwelling in the chaos of the unannounced war and
computational linguistics was “developing” only symbolically, through the separate
conceptions of language modeling written manually, at the side of the extinguished
computers.
Consequently, our project “Linguistic portrait of Georgia” did not result from the
logic and natural development of the corpus linguistics as this usually happened in
another countries (electronic textual databases/libraries + achievements of the
computational linguistics, namely successes reached in the processing of the natural
languages: conceptions of morphological and syntactic tagging, which were realized
through the invention of automatic instruments of tagging…). Our work was not
research, based on theoretical knowledge, it was rather a “instinctual reaction” of the
group of linguists on the challenge, assuming this challenge with all means available
for them.
After the fact of entering computer technologies in our lives became actual, the
idea appeared about creation large textual massive of the Georgian dialects, which
would be not only the object of studies but efficient tool for those studies. Our goal
was to process great textual massive using computer technologies and creation of new
communication area, where the model of the language, developed both in time (XX
century) and space viewpoint, and most likely to the actual situation.
We confess that we became familiar with great achievements of the corpus
linguistics in the world only after we began this work.
The Corpus of Georgian Dialects
27
2 The principles of compiling the corpus of the Georgian dialects.
2.1 The foundation principles of the main textual base
The Georgian language is represented by three Kartvelian languages and more than
two decade of their dialects. The literary Georgian itself is represented by more than
16 dialects. Three of them are distributed outside Georgia: Fereidanian (Iran, 35–40
thousand language speakers), “Turkish Georgian” (in its turn quite diverse, including
some dialectal layers), and Ingilo (Azerbaijan territory, speech of Georgians
(Christians and muslims) by origin, about 15 thousand speakers).
Scientific documentation and scientific study of the Georgian dialects began from
the very beginning of the XX century, though the early attempts of marking dialectal
entities can be seen in the ancient manuscripts (V–VII centuries). Starting from the
20s of the last century, collecting and studying of the dialectal data became one of the
priorities of the Georgian linguists scientific work. It was absolutely clear that
complex grammatical system of the Georgian could not be studied without
investigation of its dialects. The results of the dialectological studies were depicted in
the fundamental works of the Georgian linguists; for example, “Khevsurian Poetry”
collected and published by Akaki Shanidze (Khevsureti is one of the Georgian
highland ethnographic region, nowadays almost emptied of its population, hundreds
of which are scattered in different regions of the country). At the end of the 30-ies of
the last century, great Georgian linguist and historian Ivane Javakhishvili arranged
“blitz-expedition” almost in all regions of Georgia (except Iran and Turkey), to collect
the dialectal data concerning home industry and handcrafts. The data was written
down manually but strictly in accordance with the main requirements of the dialect
peculiarities. This expedition was distinguished because it was conducted in two
months and enveloped almost the whole area of the language spreading. The data of
this expedition is unique by its vast content and the possibility to represent this
content from the time and space viewpoint.
After the II World war was ended, for merely linguistic aims great campaign of
texts collecting started, which resulted in publishing the collection “Georgian
Dialectology”, in 1961 (number of pages – 600). After this separate dialectal texts
were published and then, in the 80-ies, a number of expeditions were conducted
almost in every region of Georgia; these expeditions were equipped with several
technical tools such as magnetic audio careers to fix the peculiarities of the Georgian
language in the whole area of its spreading. In the beginning of the new millennium
digital video-devices fixed hundreds of hours of the dialectal texts. Nowadays the new
texts collected by our project group are processed and integrated in the corpus
continuously, stage by stage.
We never had the “luxury” to begin our work based on the ready framework of the
electronic libraries or databases, and we could not wait till these were established, and
only then begin to think on the principles of text massive creation, on development of
28
Marine Beridze and David Nadaraia
theoretical concepts, provided that text massive creation is the indispensable precondition for development of the corpus linguistics and computational linguistics in
general (Ershov 1983).
We are building up the structure of the corpus, the principles of tagging and texts
collection all simultaneously. Furthermore, we have to carry out all kinds of work at
one time instead doing it in stages, such as: collecting the material, decoding
(discourse transcription), unification, digitalization, integration in the corpus (with
appropriate tagging) and so on.
2.2 Dialectal lexis in the corpus (database of dictionaries and raw material for
dictionaries)
As it is known from Ziph’s law, a language is a great collection of rarely
acknowledged occurrences (Kutuzov). This law was refined and approved on the
basis of the corpus linguistics. Certainly, it must always be considered when
compiling the dialectal corpus. How much we will try ‘thematically motivate” a
speaker in the process of collecting dialectal data, all the same it is impossible to
represent complete realization of the language inventory, so the perfect model of the
dialect can not be achieved. That is why we decided to integrate dialectal dictionaries
and materials for the dictionaries in the corpus, with the status of the text. This will
ensure full representation of the words in the corpus and on the other hand, it will give
additional context for realization of a lexeme (through the textual illustrations) and
besides, will assist in preparation of the cumulative dialectal dictionary. The scheme
of integration of dictionaries in the corpus is as follows: a dictionary itself is a sub
corpus, the components of which are integrated in the corpus structure with different
“status”: dictionary word-forms and the rest of the entry are related to each other as
usually is a word and its context in the concordance; as for the illustration material, it
is organized in the corpus through separate texts and joins the overall database of
words.
2.3 Illustration material of linguistic scientific literature represented in the
corpus
The dialectal lexical fund can be represented in the corpus in its complete view in the
vast textual massive but the probability of realization of absolutely completed
paradigm is very low. We suppose that linguistic “representativeness” of the dialectal
corpus must be “enhanced” by the paradigmatic descriptions of the separate dialects
(accompanied by illustrations). These are the three factors: a text, a dictionary and
data from the scientific literature, which support the creation of truly representative
dialectal corpus. Illustration data confirmed in the scientific texts will be also
integrated in the textual component of the corpus.
The Corpus of Georgian Dialects
29
3 Principles of annotation in the corpus
Annotation parameters are selected in accordance with the two requirements set to the
corpus by the working group. Those requirements consequently reveal two main tasks
of the corpus:
a. Show the most perfect model of the given linguistic area;
b. Show as complete ethno-cultural picture of the given region, as possible.
3.1 Text and meta text characteristics in the corpus are established so, that it
becomes possible to study the texts in inter and multi disciplinary ways
We consider especially important to employ these texts with the status of “oral
histories”. Such precedent has already appeared earlier and on the bases of the texts,
obtained for the dialectological purposes, we published two books about two
important stages of the Georgian history – about the Muslim population deported out
of Meskheti so called Muslim Meskhetians (Beridze, 2005) and about antigovernment rebellion in the Georgian Highland, in the period of the II World war
(Tsotsanidze, 2004).
Despite of the fact that there exist the fundamental principles of the corpus
construction, they can vary according the language peculiarities, specific historic,
cultural and linguistic characteristics. The most “variable” are characteristics of the
corpus annotation. Especially this is true to the metatextual annotation characteristics.
During the XX century, Georgia lived its life as a region of Russia (in the beginning)
and then as a Soviet Republic (till 1990). The country could not avoid historical
cataclysms that all the countries subordinated to Russia had undergone. Especially
important are the moments, which have significantly changed ethnographic, and
correspondingly the linguistic situation in Georgia. We mean some waves of
migrations – natural migrations from the highlands down to the valleys and
compulsory, repressive migrations forced on the population by the communist regime,
which nowadays are considered as occurrences of ethnographic trafficking. Recently,
the history of such events was “enriched ”with the facts of displacement of large
groups of population from Abkhazia and South Ossetia. The picture of dialects
geographical distribution greatly changed and at the end of the XX c. it differed from
that of the beginning. Many “dialectal isles” occurred on the territory of Georgia and
the country turned into a “live laboratory” for studying the migration processes and
their results. Correspondingly, the parameters of texts description in our corpus are
aimed to describe existing complex and diverse situation.
Today, the following characteristics work to describe texts in the corpus:
a) Information about narrator contains several blocks: name, ethnical and regional
origin; age, place of birth, occupation; information about parents, their origin
(to study speech of the mixed families); information about migration of the
narrator or his family: year of migration, which generation was migrated (the
30
Marine Beridze and David Nadaraia
narrator, his father, grandfather or ancestor); type of migration (compulsory,
voluntary), kind of dwelling in new places (compact or not compact).
b) Information about the text: who has recorded, who has prepared it for publishing, description of publishing (date, place) and so on;
c) Types of the text: narrative, folklore (prose), folklore (verse), fragments of
conversations;
d) Thematic characteristics;
e) Chronotypical description of the text;
f) Form of the text: printed, manuscript (from the old archives), audio, video
3.2 Morphological Tagging (most basic type of linguistic corpus annotation)
As we have mentioned above, the corpus has an ambition to represent one of the most
important segment of the Georgian language world – the model of its spatial variation
systems. Dialectal material given in the corpusal structure is approached diversely.
Some of the corpora are oriented to show interrelations between dialects and the
literary language, and thus, to show only specific characteristics of the dialect; other
corpora are focused on presentation of dialect as a whole linguistic system and the
dialectal data is given not differentiated. We are guided by the second of the named
principles, according to which, each form acknowledged in the dialect regardless
coincides it with the literary form or not, is referred to as an element of the given
dialectal system.
3.3 Primary morphological tagging
Primary morphological tagging implies initial tagging according the morphological
groups. On this stage of tagging the whole corpus material is divided into the
following clusters: noun, verb, verbal noun, invariant word-forms, non-differentiated
stem. On the first stage of morphological tagging we marked non-differentiated stems
as a separate group – these are word-forms originated from enclitic of different
elements, which only then are described according the all-constituent elements.
Primary morphological tagging process involves: description of the word-form by
the primary morphological classificator, its lemmatization, marking literary correspondence of the word-form (the next stage will be tagging by all morphological
characteristics – by classificators as well as by word-formation tools).
Introducing literary form in the tagging system will assist in:
Conducting statistic investigation:
• What should be the percentage relation between standard and variation systems
inventory in the language sub system so that it could meet minimal requirements to
exist as a separate communication model – the requirement of the differentiated
language identity;
• At which levels of hierarchy is the standard model inviolability applicable;
The Corpus of Georgian Dialects
31
• At which levels of hierarchy is the dialectal form more durable against the
influence of the literary form (or other dialects);
• At which levels of hierarchy do the “parallel” realizations occur more
frequently. To what extent can the future of the language occurrence be predicted with
help of parallel realizations.
Conducting search by different characteristics:
• Search by literary form (literary form – dialectal form in 16 sub systems);
• Search by dialectal form (dialectal form – literary form + forms of all sub
systems);
• Search by the head form (head form – all realizations of the morphological
paradigm;
• Literary head form – all realizations of the morphological paradigm (literary,
dialectical and others).
It should be noted here that lemmatization in the corpus happens only on the basis
of the acknowledged realization, as in the realism of the Georgian language it is
absolutely unacceptable to create the head form on the basis of the morphological
model (see Kryuchkova Goldyn 2009). Such categorical approach results from the
fact that in Georgian dialects one of the main “sources” of differences is the
kaleidoscopic diversity of arrangements of existing morphological models and
language inventory. That is why when conducting morphological tagging the head
form is represented by:
a. According the realizations illustrated in the text;
b. Verifying this form in the speech of the dialect career;
c. According scientific dialectological literature and dialectal dictionaries;
d. Uncertain and unspecified material should be gathered separately and specified
during the future field expeditions.
Certainly, because of some reasons, several of the word forms realizations can not
be coincide throughout almost twenty sub systems. However, equaling to the literary
head form will allow to reveal “gaps” in description of certain dialects –at the lexical,
morphemic or paradigm components levels.
System of tagging at the next stage of “building the corpus” implies the mentioned
revealing of gaps and establishing such net and interactive scientific-teaching systems
by means of that these gaps can be filed with information. In such way we want to
merge the experience in dialectology and linguistic geography in the corpus.
3.4 Possibilities of partially automatic tagging in the corpus
Being oriented on the literary form facilitates solving of two main tasks: partially
automatic tagging according the morphological group and correct correlation of the
literary language (and other language strata) and the dialectal arsenal.
By means of the morphological analizators established on purpose of the Georgian
language studies, according the literary head forms, identification of the dialectal head
form can be achieved – lemmatization and automatic annotation on the parts of
32
Marine Beridze and David Nadaraia
speech basis (parts-of-speech-tagging); this surely will entail manual annotation and
hard work of deleting occurrences of homonymy.
This technique of lemmatization can not be perfect as it applies only to that part of
lexis, the head form of which is confirmed in the text and coincides with the literary
form. But refining the process of lemmatization is possible if consider morphophonemic
changes characteristic for Georgian dialects and show relative regularity. For example,
in the Georgian dialects the stable allomorphs of vowel stems are: 0, -i, semi-vowel –i?,
also in the nouns ended in –a, we can encounter a + i > e; correspondingly, if we extend
the declension model of nouns with this additional information of vowel stems
established for the literary language, it becomes quite realistic not only to identify those
vowel stem nouns whose nominative case coincides with the literary form and their
dialectal variants as well. “patara” (small) – //patarai//patare…
Similarly, the dialectal form can be identified according the literary word form and
this will also enable us to conduct automatic tagging in reverse succession: dialect
word-form = literary word-form _ literary head form; e.g. consideration of dialectal
variations of the proverbs will provide precise identification of the verb forms, as they
differ from the literary correspondences very often only by proverbs. mo(literary)//me -//ma proverbs identification can result in the identification “ma-itana”
“me-iatana” forms with this literary variation “mo-itana” (brought) – and then in
relation of this with the literary head form “motana” (to bring). This is sufficient for
primary tagging of morphological group and for deep morphological automatic
tagging.
In 1986 “The Dictionary of Georgian Morphemes and Modal Elements” was
compiled in which dialectal variations of separate morphemes and dialectal variations
of the modal elements are registered with great precision. The morphological
inventory such as invariant words are represented with limitation as it happens in
every language and corresponding dialectal variations are given, so that the list of
such words and initial information can be delivered to the database, as a simple
identificator. The same can be said about pronouns.
For partial tagging (identification of words with their literary variants) some
phonetic rules, evidenced in the dialects with certain regularity, can be employed
successfully: e.g. voiced dental consonant “d” can give the voiceless “t” at the end of
the word. Consequently, all the words confirmed in the corpus differing only by the
final “d” and “t”, may be identified as phonetic variations of one form, e.g. adverbial
case forms: “kaca-d” (as a man) and “kaca-t”, “patarad-patarat” (as a little) and so on.
Using such technique for identification enables not only automatic lemmatization but
also primary and deep morphological tagging.
One of the main characteristics of the Georgian language is the diversification of
the dialects on the morpheme and fundamental levels. Complex and spontanic
variations of the vowels and consonants are studied in detail. Using this intellectual
resource of the corpus of dialects for automatic annotation is the future goal of our
work team.
The Corpus of Georgian Dialects
33
Summary
The experience has proved that the corpus linguistics more and more extends its
functions and purposes and has already gone beyond “utilization” and obtained far
more functions, such as ability of interdisciplinary activation, preventive purpose,
perspective of developing as a new communication system, etc. The corpus dictates
its new functions itself. Very soon the corpora created by us will independently
determine the priorities of our work, structure of the material, and so on.
The corpus of the Georgian dialects is under processing now. The main text base
still is not completely integrated in the corpus. Besides, we are facing very work - and
time - consuming process of the linguistic tagging. The multimedial part (hundreds of
hours of audio and video material) has not been included yet in the corpus.
Today we can more surely say that the corpus method of documentation and
investigation of the language data is the most effective means of maintaining language
and cultural consciousness for a small country like Georgia. Besides that, it differs
from the corpora by specially created for the small languages as it is not a static
picture, nor a catalogue of grammatical and lexical paradigms, but a ”live”
communicative environment, involving the whole language system, language world,
with all the characteristics and “genres” of the oral speech. Thus it is not an “enlarged
photo” to keep the memory about the language which is on verge of disappearing
(Kibrik 2007) but one of the means to prolog its life, its existence.
The corpus is created with the goal of documentation and investigation of the
Georgian language world, but it is constructed on the principles, which can be applied
to describe any other multi-dialectal and multi-language area.
References
[1] Kibrik A. E., Arkhipov A. V., Daniel M. A., Kodzasov S. V., Meiers,
Nakhimovski A. D. (2007). Digital processing of linguistic data for minority
languages documentation, papers of the international conference on computational linguistics, dialogue-2007.
[2] Ter-Avanesova A. V., Krilov S. A. (2006). Lexical-grammatical databases as a
tool of dialectological descriptions, papers of international conference on
computational linguistics, dialogue-2006.
[3] Kryuchkova O. U., Goldyn V. E. (2008). Textual dialect corpuses a model of
traditional rural communication, papers of the international conference on
computational linguistics, dialogue-2008.
[4] Jorbenadze B., Kobaidze M., Beridze M. (1988) Dictionary of the Georgian
morphemes and modal elements, Tbilisi.
[5] Beridze M. (2005). Direct reporting from the past (Meskhetia and Meskhetians,
1918 – 1944), Tbilisi
34
Marine Beridze and David Nadaraia
[6] Ershov A. P. Machinery fund of the Russian language – external setting. Typed text
http://ershov.iis.nsk.su/archive/eaindex.asp?lang=1&did=11377
[7] Szmercsanyi Benedikt, Hernández Nuria. (2007). Manual of information to
accompany the Freiburg Corpus og English Dialects Sampler (“FREDS”),
English department, University of Freiburg.
[8] Wood Gorodon R. (1969). Dialectology by computer, International Conference
on Computational Linguistics, Sveden.
[9] McEnery A., Wilson A. (1996). Corpus linguistics.
[10] Sinclair J. M. (1992). The automatic analysis of copora // Directions in corpus
linguistics.
[11] Zakharov V.P. (2005). Corpus Linguistics, Teaching manual.
[12] Kutuzov A.B. Course of the corpus linguistics,
http://tc.utmn.ru/files/corpus_2.pdf
Corpus of Computational Linguistics Texts
Tatiana Bobkova, Mariia Kasianenko, Kuzma Lebedev,
Valentyna Lukashevych, Pavlo Petrenko, and Liubov Grydneva
Computational Linguistics Laboratory, Kyiv National Linguistic University, Ukraine
[email protected]
Abstract. The aim of the decision was to compile a corpus of computational
linguistics texts and to study its applications in linguistics and lexicography
studies. The corpus includes the texts of handbooks, articles in English, Russian
and Ukrainian. The whole size of the Corpus is about 500 thousand word-forms
in each language. The Corpus is used in linguistics studies, for example
statistical research of functioning of words, grammar forms and collocations in
scientific texts. A software package was developed to study the texts of the
Corpus. The standard principles of text coding compilation were revised. The
glossary for searching about thousand computational linguistics terms was
designed. This tool includes an explanation of the term, its Russian and
Ukrainian translations, and contexts in all sub-corpora.
Keywords: corpus, sub-corpus, text, database, glossary, term.
1 Introduction
The aim of the research described in the article is a development of the principles of
compiling a trilingual corpus of computational linguistics texts and studying its
possible applications in linguistic researches. Similar corpora have a wide practical
use, particularly, in computational terminography, development of systems of
automatic text analysis, machine translation and information retrieval systems.
The relevance of this research is in a necessity of systematization and standardization of computational linguistic terminology, functioning in researched
professional texts.
The corpus compiled in the laboratory of computational linguistics is trilingual
and homogeneous in point of chronology, functional style and text subject.
2 Steps of corpus building
The material of the research is a collection of computational linguistic texts in
English, Ukrainian and Russian, the size of which is about 500 thousand word-forms
in each language. Compiled in the laboratory of computational linguistics of Kyiv
National Linguistic University the corpus is based on the texts of manuals, handbooks
and scientific articles of computational linguistics. For instance, the base of the subcorpus of English texts are the textbooks “The Oxford Handbook of Computational
Linguistics” [1], “Speech and Language Processing” [2], and the collected articles
36
Tatiana Bobkova et al.
“Computational Linguistics”, “Text, Speech and Language Technology”, “The Prague
Bulletin of Mathematical Linguistics” etc. The Ukrainian sub-corpus is based on the
texts of manuals “Computational linguistics” [3] and “Traditional and computational
linguistics” [4].
Authentic texts from the end of XX century till nowadays were taken for forming
each sub-corpus. The application of the trilingual material of different genres gives an
opportunity to include the vocabulary of different linguistic branches (theoretical,
structural, applied linguistics) and to represent different schools (American, European,
Kyiv, Moscow, St. Petersburg and others). The whole texts of documents are included
to the corpus that provides their structural and lexical completeness.
For compiling the corpus an additional program of addressing, which provides
texts processing by output data, was developed. Necessary data about a processed text
are input into the appropriate fields of the program. The data are the following: the
name of the text, surname and initials of the author, the name of the publishing house
and the year of edition. An annotation to this text is input separately. Then the
program enters information to the appropriate tables in the database. Coding of the
texts annotations gives an opportunity in the future to create an automatic system of
professional texts abstracting. The scheme of database of the created corpus of
computational linguistics texts is stated below.
Fig. 1. Structure of corpus database
The tables “text_table_ukr” and “main_table_ukr” are the most essential ones in
the database structure.
The table “main_table_ukr” contains texts divided into separate words, their
lemmas and grammar codes. Besides, a number of the sentence, where the word
occurred, a field with marks about its formatting and an external key-number of the
text with it are indicated for each word. The table “text_table_ukr” includes
Corpus of Computational Linguistics Texts
37
information about texts (a date of publication, a subject, information about the person
who added this text to the database, the larger text, a fragment of which is given), and
external keys for connection with the tables of the lists of the authors, publishing
houses, styles etc.
The tables “author_list_ukr”, “author_table_ukr” provide a treelike structure of
data about the authors of the texts that enables to indicate an unlimited amount of
authors.
The table “styles_table_ukr” contains information about the text style that is
input into a separate table for greater convenience.
The table “publisher_table_ukr” includes data about all publishing houses.
At this point in the corpus half-automatic morphological coding was carried out,
frequency dictionaries of word-forms and a glossary of computational linguistics
terms were compiled.
Nowadays a corpus with morphological tagging is the most widespread among
other types of corpora. Thus, coding includes not only features of belonging to the
part of speech, but also codes of grammar categories specific to this part of speech.
The result of the part of speech coding is a text containing monosemantic
morphological markers. Thus, in the majority of recent corpora programs available in
the Internet are used for the morphological coding.
In the corpus such original programs are used: 1) the program of morphological
coding of English verbs on the basis of 15 differential features [5]; 2) the program of
word search in the sub-corpus. Words can be searched by word-form, lemma and
grammar code. The creation of single system of codes for English-Ukrainian-Russian
text corpus is associated with objective difficulties that can be explained, first of all,
by different typological characteristics of the analyzed languages. It is known that in
the Ukrainian and Russian languages grammar meaning is expressed mainly
synthetically, by the means of ending, suffix, prefix, accent change, internal inflexion,
suppletive modification. On the contrary in English analytical way of grammar
meaning expression prevails (by the means of prepositions, conjunctions, articles,
auxiliary verbs, other auxiliary words and word order).
Particularly special attention should be paid to the system of English verb forms
that, in comparison with Ukrainian and Russian systems, is characterized by
considerable complexity and embranchment. To provide effective work of the corpus
correct morphological code should be given to each verb form. The program for
automatic morphological coding of verb forms carries out only initial tagging. Results
obtained after primary text processing needs manual editing and correction of the
mistakes made by the program.
Let’s observe the operation principles of the program for morphological coding of
English verb in details. For this task such additional programs are used:
1. Program for automatic definition of unambiguous verb forms - the majority of
the general list of 526 forms.
38
Tatiana Bobkova et al.
2. Program for grammar homonymy of Past Simple and Past Participle verb forms
recognition. In this case the user gets a request for manual homonymy clarification.
3. Lemmatization program that assigns lemmas to all verb forms.
The programs mentioned above are based on the processing of the database which
includes more than 1500 English verbs with their main forms. During morphological
coding the methods of context analysis are used as well.
Described programs considerably facilitate the process of morphological code
receiving for the researcher and friendly interface simplifies the process of the
correction of the mistakes made during automatic tagging.
3 Glossary of the computational linguistics terms
The final aim of creation of the trilingual corpus of computational linguistics text is
the compiling of translation terminological dictionary. That’s why the experience of
monolingual terminological dictionaries compiling in Ukraine is very useful.
Built by alphabetical principle the terminological dictionary of computational
linguistics is an explanatory one. Making up the glossary of terms was done by
combining system- and text-oriented approaches.
The register of the glossary was based on the list of English terms with their
definitions given in “The Oxford Handbook of Computational Linguistics” [1]. It is
explained by the fact that Ukrainian and Russian computational linguistics
terminology is mainly supplemented through translation, loan translation or
transliteration of English terms.
In the created glossary Ukrainian and Russian English terms matches definitions
found in the sub-corpora texts are given as well. Moreover authors refused to conform
terms definitions in different languages as it allows to reveal the differences in terms
interpretation by different linguistic schools.
To show the real functioning of terms in the glossary the examples of their use in
sub-corpora texts are given. Full sentences selected from the analyzed texts are
offered as the examples of terms use.
The glossary of English terms of computational linguistics includes about 1000
registered words. The developed software allows to search for English terms, look
through its English definition and all the examples of usage in the text massive.
Corpus of Computational Linguistics Texts
39
Fig. 2. The example of dictionary entry in the glossary of terms
To view the definition a user may type the term manually or choose it from the
list. After pushing the term its definition is given below and the equivalents of the
term in Ukrainian and Russian –on the right side. Choosing another working language
the user can look through the examples of term equivalent usage in Ukrainian or
Russian sub-corpus.
Storing of the glossary as a database allows to add new terms to the formed
register. Input and processing of new texts of the corpus automatically increase the
quantity of the terms usage examples.
Summary
Designed in the laboratory of computational linguistics of KNLU the trilingual
corpus enables future lexicographical and comparatively-typological researches while
the term glossary can be used as a reference system with a source database and as an
additional module in the machine translation system for professional literature.
40
Tatiana Bobkova et al.
References
[1] Mitkov R. (ed.) The Oxford Handbook of Computational Linguistics. Oxford:
Oxford University Press, 2003.
[2] Jurafsky D., Martin J.H. Speech and Language Processing. Upper Saddle River,
N.J.: Prentice Hall, 2000.
[3] Дарчук Н. П. Комп’ютерна лінгвістика. Київ: ВПЦ “Київський університет”, 2008.
[4] Перебийніс В. І., Сорокін В. М. Традиційна та комп’ютерна лексикографія. Київ, 2009.
[5] Morphology of English Verb: System and functioning. M.: RGDO, 2008. Р. 18-19.
[6] Штерн І.Б. Вибрані топіки та лексикон сучасної лінгвістики. К., 1998.
[7] Дерба С. М. Словник з української термінології прикладної (комп’ютерної) лінгвістики. К., 2007.
[8] Тезаурус з лінгвістичної термінології. http://www.mova.info
“We Only Say We Are Certain When We Are Not”:
A Corpus-Based Study of Epistemic Stance*
Vaclav Brezina
University of Auckland, New Zealand
[email protected]
Abstract. This paper investigates epistemic stance in speech using two
corpora of spoken English: the spoken part of the BNC (British National
Corpus) and MICASE (Michigan Corpus of Academic Spoken English). In
particular, the paper uses corpus evidence to evaluate Halliday’s (1994)
famous dictum about epistemic stance marking: “The importance of modal
features in the grammar of interpersonal exchanges lies in an apparent
paradox on which the entire system rests – the fact that we only say we are
certain when we are not” (p. 362). The analysis of four epistemic stance
markers (must, certain, sure and certainly) brings evidence in support of
Halliday’s claim. It appears, however, that this claim is valid only if
uncertainty is understood in intersubjective rather than in subjective
(psychological) terms.
1 Introduction
In every statement we make, we put across not only some propositional content, but
we also indicate how certain or uncertain we are about it (i.e. we indicate our
epistemic stance). By indicating our epistemic stance, we simultaneously evaluate
the proposition, position ourselves and align with the hearer(s) (Du Bois 2007).
Epistemic stance therefore plays a crucial role in any linguistic interaction and
spoken interaction in particular.
It is important to realise that epistemic stance can be either explicitly marked (by
gestures, intonation or purely linguistic means) or implied by the pragmatics of the
speech act. From the perspective of pragmatics, if a speaker makes an unmitigated
statement, she commits herself to the truth of the proposition and can be held
accountable for it (Searle1985, Holmes 1984a). A simple statement of the type this
is the case can thus be understood as the neutral way of expressing the speaker’s
certainty about the propositional content. We can therefore asks the question: why
do speakers choose to mark their certainty linguistically forming utterances of the
type this certainly is the case or this must be the case?
A remarkable answer to this question is offered by Halliday (1994), who claims
that:
*
I would like to thank Dana Gablasova for her invaluable comments on an earlier version of
this paper.
42
Vaclav Brezina
[t]he importance of modal features [i.e. epistemic stance] in the grammar of
interpersonal exchanges lies in an apparent paradox on which the entire system rests –
the fact that we only say we are certain when we are not (p. 362).
The aim of this paper is to evaluate Halliday’s famous dictum using evidence
from two corpora of spoken English: the spoken part of the BNC and MICASE1. The
former represents the spoken British English from the early 1990s, while the latter
captures spoken academic English (American variety) at the turn of the millennia. In
short, this paper uses corpus evidence to investigate speakers’ choices of expressing
certainty in spoken language.
2 Certainty and uncertainty in spoken language
In literature, speaker’s certainty and uncertainty about the propositional content of
an utterance has been treated under different labels:modality (Coates 1987, 2003;
Palmer 2001; Nuyts 2001, 2006; DeLancey 2001), modal evaluation (Thompson &
Hunston 2001), intensity (Labov 1984), evidentiality (Chafe 1986), hedging (Holmes
1984b; Hyland 1996), and epistemic or epistemological stance (Barton 1993; Biber
et al. 1999; Kärkkäinen 2003; Biber 2006).
Each of these approaches stresses a different aspect of epistemicity. Drawing on
the previous research, the present study proposes a comprehensive framework for
analysing epistemic stance in spoken language (see Table 1).
Following DeLancey (2001), it can be argued that the function of epistemic
stance is to qualify a proposition in relation to ideal knowledge status. There are
three aspects of this status:
•
•
•
What is said is known by the speaker by direct experience.
What is said is assumed to be certainly true.
What is said is fully consistent with the rest of the speaker’s knowledge of the
world.
It is important to realise that a unique feature of any epistemic qualification is
the fact that unlike speaker’s comments on the style of speaking and her evaluative
attitude, epistemicity is inherently present in every utterance (although not always
linguistically marked). Every statement in language can therefore be said to have
the following underlying structure: epistemic qualification [proposition]
This underlying structure is reflected in the type of speech act (pragmatics of
epistemicity) as well as in the linguistic structures which mark the speaker’s
(un)certainty (lexico-grammatical epistemicity marking) – see Table 1. As is
apparent from Table 1, linguistic marking of (un)certainty involves different word
classes such as adverbs, adjectives, nouns, modal auxiliaries and lexical verbs,
which appear in various epistemic structures .
1
British national corpus and Michigan corpus of academic spoken English.
A Corpus-Based Study of the Epistemic Stance
43
underlying structure (semantics)
epistemic qualification [proposition]
patterns (grammar & pragmatics of epistemic stance)
patterns
BNC examples
I. (pragmatic aspect of epistemicity) [proposition]
unmarked/unmodlised/unmodified illocutionary force
1. Statement (Searle’s
S1 [ ??? ] alcoholic.
assertive speech act)
S2 [ ??? ] ninety nine and a half per cent.
S1 [laugh] , erm I'd imagine it's fairly potent.
certainty
S2 Yes, well I'm telling you it is[ ??? ].
(explicit, with a verbum dicendi)
2. Question (Searle’s
directive speech act)
uncertainty
It is eight o'clock though [laugh] (implicit)
S1: When did you watch Lady And The Tramp?
S2: In the lounge.
S1: I know where you watched it, I'm asking you when you
watched it.
(explicit, with a verbum dicendi)
When did you go and get that? (implicit)
II. (lexico-grammatical marking of epistemicity) [proposition]
marked
3. Tag questions
He's got a real little face hasn't he?
I think that's part of the reason why she said you've got
your hands full.
They seem to use them a lot, buses, don't they?
4. Epistemic matrix clause
[epistemic matrix clause][subordinate clause] It's pretty certain that you can stay.
It's not likely to be over I've come to the conclusion that he, I mean, his work is, it
was absolutely superb...
5. Epistemic parentheticals, comment
...but they're giving them a packed lunch so I believe.
clauses, epistemic afterthoughts
...and the bloke I remembered best of all, I think, was oh
[main clause] [epistemic parenthetical]
Cakey
6. Epistemic adverbials
They definitely had knocked off somebody else..
[main clause[epistemic adverb]]
7. Epistemic modals
They must have found it.
[main [modal] clause]
III. Combinations of lexico-grammatical markers
8. Congruous
So I think probably that's what it was, but don't know.
9. Incongruous
...I think definitely I think it must be the private scheme...
Table 1. Epistemic structures
Epistemic markers (linguistic signals of epistemic stance) play an important part
in the process of negotiating knowledge. Knowledge has been traditionally
understood as justified true belief (Noddings 2007). Epistemic markers, therefore,
first of all, indicate how likely it is that a particular proposition is true. (This is at
least one of the functions of these multifunctional items.)
Second, epistemic markers in spoken discourse interaction represent the means
of justification of one’s beliefs. The justification process, however, does not take
place in a vacuum, but in a very complex social setting, which reflects various social
power relations. When analysing epistemic patterns, we therefore have to take into
consideration the following socio-pragmatic aspects:
44
•
•
•
Vaclav Brezina
If a speaker claims something she commits herself to the truth of the proposition
and can be held accountable for it. In this respect, the degree of certainty can be
understood as the degree of commitment to the truth of a particular proposition.
If a speaker says something with a high degree of certainty, she may find herself
in a position of contradicting what another person has said. Hence, this may be
potentially face-threatening.
There exist numerous (often implicit) rules and social norms given by the
cultural/social as well as the individual context.
3 Data and methodology
The corpora used in this research (BNC-spoken part, MICASE) have been chosen to
represent general spoken and spoken academic English. From each corpus a subcorpus of highly interactive language (dialogue) has been extracted so that we can
get a better insight into the dynamic process of knowledge negotiation. The details
about these sub-corpora (BNC-CONV and MICASE-INT) can be found in Table 2
below.
Corpora
Tokens
No. of
Speaker’s
speakers gender
33% male
BNC4,233,955 1525 (?) 37% female
CONV
30% unkn.
48% male, 52%
MICASE-INT 564,683 520
female
Genre
Discourse Variety Period
mode
of English
informal
conversation
highly
UK
interactive
spoken academic highly
USA
interaction
interactive
early 1990s
1998–2001
Table 2. Characteristics of the corpora
BNC-CONV2, the larger of the corpora, represents the British variety of spoken
English from the early 1990s. It contains more than 4.2 million words of transcribed
speech (Aston and Burnard 1998, Crowdy 1995).
MICASE-INT, consists of more than half a million words of transcribed academic
spoken English (American variety) from various university contexts (lectures,
seminars, lab meetings, student presentations, office hours etc.) – see MICASE
Manual 2007.
As is clear from the previous discussion (see Table 1 for summary), there are
copious lexico-grammatical means, which the speakers can employ to mark
certainty in speech. The large epistemic lexico-grammatical structures are, however,
relatively difficult to search for in the corpus of spoken language as there exist
numerous variants and modifications of these structures (see examples [1] – [4]
below).
[1] Certainly has.
[2] They're certainly not quick.
[3] Cos I, I certainly didn't say nothing to her.
[4] Not, oh it was bad, it were it were, well, well stood up all traffic certainly.
(BNC-CONV)
2
BNC-CONV corresponds to the demographic part of the BNC-spoken sub-corpus.
A Corpus-Based Study of the Epistemic Stance
45
Nevertheless, most of these structures contain an epistemic key word (e.g.
certain, conclusion, certainly etc.) which can be easily searched for using standard
concordancing software (e.g. Xaira, MonoConc). I will call these keywords
epistemic candidates. An extensive list of epistemic candidates can be found in
Biber (2006) and is summarised in Table 3. The epistemic candidates are further
divided according to the degree of certainty which they mark and each of the high
certainty markers is accompanied by a raw frequency count based on the BNCCONV corpus.
High certainty (raw frequencies in Lower certainty (likelihood)
BNC-CONV)
actually (3310), always (2813), certainly
apparently, evidently, in most cases,
ADVERBS
ADJECTIVES
NOUNS
LEXICAL
VERBS
MODALS
(426), definitely (503), indeed (207),
inevitably (1), in fact (427), never (4113), of
course (1217), obviously (672), really (9128),
undoubtedly (5), without doubt (0), no doubt
(60)
apparent (5), certain (288), clear (316),
confident (25), convinced (32), correct (97),
evident (2), false (55), impossible (43),
inevitable (3), obvious (84), positive (33),
right (14984), sure (1958), true (726), wellknown (0)
assertion (0), conclusion (21), conviction (3),
discovery (5), fact (323), knowledge (40),
observation (6), principle (12), realization (1),
result (32), statement (34)
conclude (4), demonstrate (11), determine
(19), discover (55), find, (2744) know
(33526), learn (446), mean (12351), notice
(501), observe (7), prove (78), realis/ze
(545), recognis/ze (92), remember (2619), see
(16527), show (1147), understand (565)
necessity
in most instances, kind of, maybe,
perhaps, possibly, predictably,
probably, roughly, sort of
doubtful, likely, possible, probable,
unlikely
assumption, belief, claim, contention,
feeling, hypothesis, idea, implication,
impression, notion, opinion,
possibility, presumption, suggestion
appear, assume, believe, consider,
doubt, expect, find, forget, gather,
guess, happen, hypothesize, imagine,
judge, know, learn, predict, presume,
presuppose, pretend, reckon,
remember, seem, speculate, suppose,
suspect, tend, think
prediction
must (3014), should (4396), (had) will, would, shall,
better, have to, got to, ought to
be going to
(454)
possibility
can, could, may, might
Table 3. Epistemic candidates according to Biber (2006), Biber et al. (1999)
It is important to bear in mind that not all epistemic candidates function as
markers of the speaker’s epistemic stance in all contexts, as is evident from
examples [5] and [6].
[5] it only acts this way in a certain circumstance... (MICASE-INT)
[6] he was in there and he said my old man knows for certain that at that time they
were under fire (BNC-CONV)
In example [5] the adjective certain is synonymous with the adjective particular
and therefore does not contribute to the marking of the speaker’s epistemic stance
(see 4.2). In example [6], on the other hand, certain marks an epistemic stance.
However, this does not represent the speaker’s epistemic stance, but a reported
46
Vaclav Brezina
epistemic stance. Instances such as [5] and [6] have been excluded from further
analysis as irrelevant. In practice, this was done by manually checking all
concordance lines with epistemic candidates.
For the purposes of the present study, three epistemic candidates have been chosen:
the modal auxiliary must, the adjective certain and the adverb certainly. These represent
a variety of lexico-grammatical options available to speakers for marking certainty.
Nevertheless, these forms by no means exhaust all the possibilities of marking high
degree of epistemic stance (see Table 2). This limitation, therefore, needs to be kept in
the back of our mind when we look at and interpret the data.
4 Analysis
4.1 Must
The modal auxiliary must is one of the most frequent epistemic candidates. In fact,
as previous corpus-based research shows (Holmes 1982, Biber et. al. 1999) modal
auxiliaries are the most common linguistic markers of epistemic stance. It is also
primarily the epistemic must that Halliday has in mind when he makes the claim
about the paradox inherent in the epistemic system (see Halliday 1994: 354ff). It is
therefore more than justified to start the discussion with this form.
The modal must can be either epistemic (as in example [7]), or deontic (as in
example [8]). Only the epistemic uses of must will, however, be a subject of further
analysis as only these are relevant to certainty marking.
[7] Erm it must have been the last day, it must have been Friday... (BNC-CONV)
[8] must try a lot harder. (BNC-CONV)
Corpora
BNC-CONV
MICASE-INT
"must"3
AF
3014
91
NF
711.9
161.2
Epist. stance
AF
1929
54
NF
456
95.6
%
64
59.3
EV
195
21
UN
251
9
Table 4. Frequency distribution of must
The data show that marking epistemic stance is the predominant function of
must both in informal (64 per cent) and academic (59.3 per cent) spoken interaction.
This is in accordance with Biber et. al’s (1999: 494) finding that face-to-face
interaction favours epistemic must over deontic must since in face-to-face
interaction, obligation is expressed more obliquely (i.e. without employing the
deontic must and thus in a less face-threatening way).
3
The columns with the heading “must” report the overall frequency of the form must (i.e.
must as an epistemic candidate) in the corpora, whereas the columns with the heading
“Epist. stance” show the frequency distribution of must as a marker of the speaker’s
epistemic stance.
AF ... absolute frequency; NF... frequency normalised to the basis of one million; EV...
cases of epistemic stance, which are explicitly evidential;UN...cases, in which epistemic
must occurs in the context of marked uncertainty (i.e. with expressions such as maybe,
I think etc.).
A Corpus-Based Study of the Epistemic Stance
47
Traditionally (Palmer 1990, 2001), epistemic must has been seen as a marker
which appears in situations in which the speaker makes judgement on the basis of
direct evidence (observation). The corpus data partly confirm this claim showing
that indeed epistemic must primarily appears in evidential contexts, which are,
however, not necessarily based on direct evidence. Cf. example [9], in which the use
of must is triggered by general experience or “folk” knowledge.
[9] A: I had a very strange dream!
B: Must have been something we’d eaten. (BNC-CONV)
It is interesting to notice that many of these evidential contexts are explicitly
evidential (195 examples in BNC-CONV and 21 examples in MICASE-INT).In
theses cases, speakers either provide reasons for their claims (example [10]) or use
a deictic reference justifying the claim (example [11])
[10] hm, yeah i, i (don’t know) Rasmussen must have been easy or something cuz,
it wasn’t too hard for, for us (MICASE-INT)
[11] A: Was there an accident? I dunno. This is an ambulance here.
B: An ambulance. so it must be an accident! Look like it doesn’t it? (BNCCONV)
The evidential aspect of the contexts in which must appears provides an
important clue for evaluating the degree of certainty connected with the use of the
epistemic must. Although providing reasons for one's claims may strengthen these
claims, it at the same time brings about a certain degree of tentativeness. The mere
fact that the speaker opts for justification of her statement suggests that in that
particular situation the statement is not perceived as obvious and may be
challenged.
It is important to bear in mind that it is the pragmatics of the speech act that
provides the basic epistemic ground (see section 2). This means that many of our
statements are perceived as certain without being linguistically marked for certainty.
A simple unmarked statement can therefore be interpreted as unproblematically
certain. The speaker's choice to mark the statement using an epistemic marker such
as must adds another layer of epistemic meaning.
A further insight into this process can be gained by looking at examples, in
which an epistemically marked statement appears side by side the same statement,
which is unmarked (see examples [12] and [13]).
[12] Yes, it was taken about July that was, must have been. Because I remember...
when Tom was on holiday... tha - that ground was all ... dry wasn'tit? (BNCCONV)
[13] he tipped out a pile of library tickets onto the thing and it was, it was a, it
was, must of been, it was hundreds...(BNC-CONV)
In both examples the speaker moves between an unmarked statement and a
statement qualified by the epistemic must. In example [12], the speaker first makes
a statementit was taken about July, which itself is rather tentative (since the speaker
employs the approximator about). What follows is a process of justification (in this
48
Vaclav Brezina
case taking place within a single turn), in whichthe speaker provides arguments for
her original statement. [13] is then an example of how a speaker moves back and
forth between the unmarked and the marked statement in the process of thinking.
These examples suggest that epistemic must is employed in situations in which the
speaker’s statement is far from being unproblematically certain.
Moreover, the data indicatethat epistemic must occurs relatively frequently (251
examples in BNC-CONV and 9 examples in MICASE-INT) in the context of
explicitly marked uncertainty as in examples [14] and [15].
[14] He must have gone out. I dunno.. (BNC-CONV)
[15] ...i mean i think that’s something that must be sort of, maybe the U-S gets of
that to some degree... (MICASE-INT)
4.2 Certain (and sure)
The adjective certain is another epistemic candidate which will be discussed in
relation to Halliday’s claim. In fact, the structure I’m certain (that)... is the most
explicit expression of speaker’s high degree of certainty about a proposition. It
closely reflects the underlying epistemic pattern (see section 2) and there is a clear
attribution of the epistemic stance to the speaker.
There are however, two important issues which need to be considered in relation
to this epistemic candidate. First, there are two basic uses of the adjective certain,
which are reflected in the corpora and which are found in most English dictionaries
(e.g. COBUILD, CALD, OALD, OED4). Certain can indicate either speaker’s
certainty about something (epistemic stance) as in [16] or it can be used to refer to a
particular thing without specifying this item as in example [17].
[16] Oh I'm certain it is. (BNC-CONV)
[17] i think ideally they wanted you to write poems on a certain theme and put
‘em into a book... (MICASE-INT)
It is interesting to notice that both in ordinary conversation as well as in
academic interactions the latter use of the form certain (i.e. certain meaning
particular) is much more frequent than the former use (i.e. epistemic stance) -see
Table 5. In fact, there is no attestation of the adjective certain expressing the
speaker's epistemic stance in MICASE-INT and only 17 instances of it in BNCCONV. How can this fact be accounted for? This question leads us to the second
consideration.
4
Collins COBUILD Advanced Learner's English Dictionary, Cambridge Advanced Learner's
Dictionary, Oxford Advanced Learner’s Dictionary, Oxford English dictionary.
A Corpus-Based Study of the Epistemic Stance
Corpora
BNC-CONV
MICASE-INT
“certain”5
AF
NF
288
68
124
219.6
AF
17
0
49
Epist. stance
NF
%
4
6
0
0
Table 5. Frequency distribution of certain
The data show that speakers are rarely as explicit as to use the structure I’m
certain (that)... Nevertheless, speakers have a choice to use a similar structure I’m
sure (that)..., which is less formal.As is clear from Table 3, the adjective sure is
considerably more frequent in casual conversation (BNC-CONV) than the adjective
certain. The phrase I'm sure appears 563 times (133 per million words) in the BNCCONV corpus and 45 times (80 per million words) in MICASE-INT.
There are two major functions of the phrase I am sure (that)... First, in informal
conversation as well as in the academic interactions, the phrase is often connected
with expressing encouragement as in example [18]. Here the social function is more
dominant than the epistemic one.
[18] A: Well, the first one's [job] always gonna be the worst one!
B:Mm! Certainly is! But I say, I'm sure your capable of taking the bull by the
horns. (BNC-CONV)
Second, the phrase appears as a linguistic marker of epistemic stance proper.
When I’m sure is employed with this function, this is very often in situations in
which the certainty of a statement is in dispute: I’m sure is thus often contrasted
with I’m not sure or I don’t know as in examples [19] and [20]. It may also appear
in situations, in which the certainty of a claim is challenged by the addressee
(example [21]).
[19] A:...it's either Iceflow or Iceland I'm not sure which one it is
B: I'm sure it was Iceland but that in King Street is Iceflow in't it? (BNCCONV)
[20]
agree also because of the fact that children will see this, and i don't know
i'm sure parents wouldn't want them to see this or for them to emulate this
behavior (BNC-CONV)
[21] A: she's wrong.
B:Mm?
A: I'm sure Richard said Isle of Man. She keeps saying, Isle of Wight Isle of
Wight.. (BNC-CONV)
5
The columns with the heading “certain” report the overall frequency of the form certain
(i.e. certain as an epistemic candidate) in the corpora, whereas the columns with the
heading “Epist. stance” show the frequency distribution of certain as a marker of the
speaker’s epistemic stance.
AF ... absolute frequency; NF... frequency normalised to the basis of one million.
50
Vaclav Brezina
4.3 Certainly
The epistemic adverb certainly, as Simon-Vandenbergen and Aijmer (2007) point
out, can be considered “a prototypical adverb in the field of modal certainty” (p.
85). Although it is not as common as the adverbial epistemic candidates of high
certainty degree such as of course, really, actually or always (see Table 3), it comes
immediately to one’s mind, when one starts considering the various ways in which
we can say that we are certain.
This does not mean that certainly is used exclusively to express the speaker’s
epistemic stance. In spoken discourse, it is sometimes used as a positive response to
a request or a question as in [22] and [23].
[22] A: Please may I have some tea?
B: Certainly my dear. (BNC-CONV)
[23] A: can I leave it open?
B: Mhm, certainly. (BNC-CONV)
Nevertheless, as is clear from Table 6, the predominant function of certainly is
to mark the speaker’s epistemic stance (89 per cent of cases in BNC-CONV and 100
per cent of cases in MICASE-INT). The question to ask at this point is: What are the
typical contexts in which certainly appears?
Corpora
BNC-CONV
MICASE-INT
"certainly"6
AF
426
49
NF
100.6
86.8
AF
379
49
NF
89.5
86.8
Epist. stance
%
89
100
CON
79
14
UN
46
11
Table 6. Frequency distribution of certainly
The data indicate that certainly often appears in a combination with an epistemic
marker with a lower degree of certainty such as I think, almost or a tag question
(see examples [24] and [25]).
[24] um, i think that’s, certainly, can be useful and ... (MICASE-INT)
[25] They’re certainly not as old as ours are they? (BNC-CONV)
In these cases, the speaker undermines a strong epistemic claim by a lower
certainty marker. This can be explained by reference to the dynamics of the spoken
interaction, in which the speaker often seeks an approval from the addressee in the
process of knowledge negotiation.The incongruous combination of two epistemic
markers thus leaves more space for the addressee to disagree with the speaker
6
The columns with the heading “certainly” report the overall frequency of the form
certainly (i.e. certainly as an epistemic candidate) in the corpora, whereas the columns
with the heading “Epist. stance” show the frequency distribution of certainly as a marker
of the speaker’s epistemic stance. AF ... absolute frequency; NF... frequency normalised to
the basis of one million; CON...the contrastive use of certainly;UN...cases, in which
epistemic certainly occurs in the context of marked uncertainty (i.e. with expressions such
as I think etc.).
A Corpus-Based Study of the Epistemic Stance
51
without threatening the speaker’s face. These combinations can therefore be said to
express a certain kind of socially motivated uncertainty.
However, even if we look solely at instances, in which certainly appears as the
only epistemic marker, we can very often find this adverb in the context of doubt
(79 examples in BNC-CONV and 14 examples in MICASE-INT). As SimonVandenbergen and Aijmer (2007: 96) point out, “certainly often functions in a
context of contrast with uncertainty.” This can be shown on the following examples:
[26] related to the weather or not i don’t know. but certainly with it being darker
when the when the the you know the cloud... (MICASE-INT)
[27] Anyway, call it south London. It’s certainly not a north London accent.
(BNC-CONV)
[28] she was buying a house for whatever it was it certainly was a lot of money.
(BNC-CONV)
[29] Certainly not in my age group. (BNC-CONV)
In all the examples above, the speakers use certainly in a situation, in which they
lack the relevant precise knowledge. In these contexts, certainly introduces a piece
of information, which is usually vague or not as relevant, but speakers can be
relatively certain about it (in contrast to the piece of information which they do not
know). The prototypical structure in these cases is I don’t know this, but certainly....
5 Conclusion
We have reached a point at which we can evaluate Halliday’s claim about the
markers of certainty in English. Three general points which emerged from the data
are worth emphasising at this stage.
First, all of the certainty markers in this study (must, certain/sure and certainly)
play an important role in the process of knowledge negotiation. In the context of
spoken face-to-face interaction (be it informal conversation or academic
interaction), any piece of information which may not be obvious either to the
speaker or the hearer is likely to be epistemically marked. Although in a situation
like this, one would expect an epistemic marker of lower degree of certainty (such
as maybe or possibly) to be used, it is often the case that must, certain/sure and
certainly appear in this context.
Second, these epistemic markers are closely connected with uncertainty. The
uncertainty may be present in the immediate context as in the incongruous
combinations such as I think it must, I don’t know I’m sure, I think it certainly etc. or
it can be implied by the need of justification of the speaker’s statements (the
evidentiality context of must). Sometimes also (as is often the case with the adverb
certainly) the uncertainty about one piece of information is contrasted with certainty
52
Vaclav Brezina
about another piece of information which, however, is not as precise or relevant the
original piece of information.
Third, the notion of uncertainty which appropriately captures the dynamics of
the spoken interactions is not the subjective (psychological) uncertainty, but rather
intersubjective uncertainty, which reflects the process of knowledge negotiation and
knowledge co-construction in dialogue. As linguists, we do not claim to have a
direct access to the processes that take place in the speaker’s mind. For this reason,
we cannot say much about the speaker’s subjective uncertainty. However, thanks to
corpora, we do have an access to the typical epistemic patterns, in which, as the data
indicate, markers of certainty and uncertainty often occur side by side.
To conclude, we can say that Halliday’s claim is to a large extent justified, if
uncertainty is understood as intersubjective uncertainty. In this light, Halliday’s
claim can be paraphrased as: We only say we are certain when there is a need to
negotiate the validity of what we say.
References
Aston, Guy, and Lou Burnard. 1998. The BNC handbook: Exploring the British
national corpus with Sara. Edinburgh: Edinburgh University Press. Barton, Ellen
L. 1993. Evidentials, argumentation, and epistemological stance. College
English 55 (7):745-769.
Biber, Douglas. 2006. University language: A corpus-based study of spoken and
written registers. Amsterdam: John Benjamins.
Biber, Douglas, Stig Johansson, Geoffrey Leech, Susan Conrad and Edward
Finegan. 1999. Longman grammar of spoken and written English. Harlow:
Longman.
Chafe, Wallace. 1986. Evidentiality in English conversation and academic writing.
In Evidentiality: The linguistic coding of epistemology, edited by W. Chafe and
J. Nichols. Norwood: Ablex.
Coates, Jennifer. 1987. Epistemic modality and spoken discourse. Transactions of
the Philological society 85 (1):110-131. 2003. The role of epistemic modality in
women’s talk. In Modality in contemporary English, edited by R. Facchinetti, M.
Krug and F. Palmer. Berlin: Walter de Gruyter.
Crowdy, Steve. 1995. The BNC spoken corpus. In Spoken English on Computer,
edited by G. Leech, G. Myers and J. Thomas. Harlow: Longman.
DeLancey, Scott. 2001. The mirative and evidentiality. Journal of pragmatics 33
(3):369-382.
Du Bois, John W. 2007. The stance triangle. In Stancetaking in discourse:
subjectivity, evaluation, interaction, edited by R. Englebretson. Philadelphia:
John Benjamins.
Halliday, M. A. K. 1994. An Introduction to functional grammar London: Edward
Arnold.
A Corpus-Based Study of the Epistemic Stance
53
Holmes, Janet. 1982. Expressing Doubt and Certainty in English. RELC Journal 13
(2):9-28. 1984a. Modifying illocutionary force. Journal of Pragmatics 8 (3):345365. 1984b. Hedging your bets and sitting on the fence: Some evidence for
hedges as support structures. Te Reo 27 (1):47-62.
Hyland, Ken. 1996. Talking to the academy: Forms of hedging in science research
articles. Written communication 13 (2):251-281.
Kärkkäinen, Elise. 2003. Epistemic stance in English conversation: A description of
its interactional functions, with a focus on I think: John Benjamins. 2007. The
role of I guess in conversational stancetaking. In Stancetaking in discourse:
subjectivity, evaluation, interaction, edited by R. Englebretson. Philadelphia:
John Benjamins.
Labov, William. 1984. Intensity. In Meaning, form, and use in context: Linguistic
applications, edited by D. Schiffrin. Washington, D.C: Georgetown University
Press.
MICASE Manual: The Michigan Corpus of Spoken Academic English. 2007. Place
Published: English Language Institute, The University of Michigan.
http://lw.lsa.umich.edu/eli/micase/MICASE_MANUAL.pdf
(accessed July 2009).
Noddings, Nel. 2007. Philosophy of education. Boulder: Westview Press.
Nuyts, Jan. 2001. Epistemic modality, language, and conceptualization: A
cognitive-pragmatic perspective. Amsterdam: John Benjamins. 2006. Modality:
overview and linguistic issues. In The expression of modality, edited by W.
Frawley. Berlin: Mouton de Gruyter.
Palmer, Frank Robert. 2001. Mood and modality. Cambridge: Cambridge
University Press. 1990. Modality and the English modals. London: Longman.
Searle, John R. 1985. Expression and meaning: Studies in the theory of speech acts.
Cambridge: Cambridge University Press.
Simon-Vandenbergen, Anne-Marie, and Karin Aijmer. 2007. The Semantic field of
modal certainty. A corpus-based study of English adverbs. Berlin: Mouton de
Gruyter.
Thompson, Geoff, and Susan Hunston. 2000. Evaluation: An introduction. In
Evaluation in text: Authorial stance and the construction of discourse, edited by
S. Hunston and G. Thompson. Oxford: Oxford University Press.
A Model for Corpus-Driven Exploration and
Presentation of Multi-Word Expressions
Annelen Brunner and Kathrin Steyer
Institute for the German Language, Mannheim, Germany
Abstract. In this paper we outline our corpus-driven approach on detecting, describing and presenting multi-word expressions (MWE). We make use of large
corpora and statistical data to explore and visualize the rich interrelations, patterns and types of variances of MWE. In order to do this, we have developed a
method of linguistically interpreting collocational data in a structured way (cf.
[Steyer/Brunner 2009]). Several levels of abstraction build on each other: surface
patterns, Lexical realizations (LR), MWE and MWE patterns. Generalizations are
made in a controlled way and in adherence to corpus evidence. The method helps
to identify and describe MWE in a way that gives credit to their flexible nature
and their role in language use.
1
Methodological foundations
Our approach is corpus-driven as defined by Tognini-Bonelli who states:
“In a corpus-driven approach the commitment of the linguist is to the integrity
of the data as a whole, and descriptions aim to be comprehensive with respect
to corpus evidence.” [Tognini-Bonelli 2001, p. 84]
Following this basic principle, we work empirically with large quantities of corpus data
and generate our linguistic hypotheses and generalizations bottom up. The following
steps are crucial to our interpretative practice (cf. [Steyer/Lauer 2007, p. 494]):
– Study of all evidence of the corpus and acceptance of this evidence: We use collocation profiles as well as pattern matching to get a starting point for our analysis
that is as close to real life usage of language and as objective as possible.
– Generation of hypotheses on the basis of the evidence: We take interactive steps in
formulating and refining pattern matching queries to study the evidence.
– Empirical checking of those hypotheses: We check the results of our queries for
plausibility and revise if necessary.
– Generalization leads to usage rules: In our model, generalization happens on several
hierarchical levels and is detailed by narrative comments if necessary. Usage is
always the key factor for justifying generalization.
We have a broad concept of MWE, which is heavily influenced by experience
with empirical language data and centers around usage. Our German label – Usuelle
Wortverbindungen (first used in [Steyer 2000]) – can be translated as ‘MWE which are
common in usage’. We adhere to Firth’s contextual theory of meaning, here summarised
by Tognini-Bonelli:
Corpus-Driven Exploration of MWE
55
“In the Firthian framework the typical cannot be severed from actual usage,
and ‘repeated events’ are the central evidence of what people do, how language
functions and what language is about.” [Tognini-Bonelli 2001, p. 89]
Following this idea, we regard MWE as conventionalized patterns of language
use that manifest themselves in recurrent syntagmatic structures and have acquired
a distinct function in communication (cf. [Feilke 2004]; cf. [Brunner/Steyer 2007],
[Steyer/Brunner 2009]). They must have at least two concrete lexical components, but
may also contain abstract components representing a certain subset of lexical items or
even a general grammatical class. Neither idiomaticity nor idiosyncrasy on a grammatical or lexical level is a necessary criterion for MWE in our definition. MWE can have
a perfectly regular structure, as long as they work as functional chunks in language use.
Our approach to analysis is similar to that of Hanks detailed in the description of
his Corpus Pattern Analysis (CPA):
“Concordance lines are grouped into semantically motivated syntagmatic patterns. Associating a ‘meaning’ with each pattern is a secondary step, carried out
in close coordination with the assignment of concordance lines to patterns. The
identification of a syntagmatic pattern is not an automatic procedure: it calls for
a great deal of lexicographic art. Among the most difficult of all lexicographic
decisions is the selection of an appropriate level of generalization on the basis
of which senses are to be distinguished.” [CPA]
CPA aims at describing single words (cf. also [Hanks 2008]), while we are interested
in MWE, which adds an additional level of complexity as identifying the surface form
itself requires an interpretative effort. To handle the difficulties of generalization, our
model has several hierarchical levels, which will be presented below.
2
2.1
Model of analysis
Prerequisites
The basis of our work is collocation profiles, computed from a very large corpus of
written German, DeReKo (Deutsches Referenzkorpus), which currently consists of over
three billion tokens (cf. [KLa]). The sophisticated method used for generating these
profiles was developed by Cyril Belica ([Belica 1995]). It takes a target word form as
input and computes the word forms that appear in the vicinity of this target word more
often than statistically expected. These partner word forms are clustered on multiple
levels. The KWIC (key word in context) lines from the corpus are grouped into collocation clusters, according to the word forms they contain. The total of all clusters
generated for a target word form is called its collocation profile. Collocation analysis is available for the IDS corpora via the COSMAS II corpus research tool (http:
//www.ids-mannheim.de/cosmas2) and can be customized in various ways. For details
see [KLb].
Starting point for our study of MWE is the collocation profile of a target word
form. Figure 1 shows a snapshot from such a profile for the word form Ohren [ears].
Though collocation analysis offers lemmatization, we do not use this setting, neither
56
Annelen Brunner et al.
Fig. 1. Collocation profile of Ohren (computed 29 July 2009 via COSMAS web)
for the target word form nor when computing its collocates. This is because empirical
research shows that the behavior and contexts of different realizations of a lemma are
often quite different. These distinctions would be obfuscated by lemmatization. Also,
lemmatization of a word form is already an abstraction and we want to be very careful
not to make assumptions. In this respect, we follow Sinclair who pointed out:
“There is a good case for arguing that each distinct form is potentially a unique
lexical unit, and that forms should only be conflated into lemmas when their
environments show a certain amount and type of similarity.” [Sinclair 1991,
p. 8]
Consequently, we study the profiles of several word forms which belong to the same
lemma separately and make generalizations only at a much later stage of analysis.
We thus start with a collection of KWIC lines which contain the non-lemmatized
target word form (e.g. Ohren [ears]) and are grouped according to the collocates that
have been identified for this target word. This gives us a very good starting point, as the
statistical method has already detected regularities within the data with extremely few
a priori assumptions. Our goal now is to make use of the information given by corpus
data and statistical analysis in a structured and controlled way. We have designed four
different levels of abstraction:
–
–
–
–
The level of surface patterns
The level of Lexical realizations (LR)
The level of MWE
The level of MWE patterns
Each of these levels builds on the previous and on each level we work manually when
grouping, correlating and commenting on the phenomena we observe. The model is
Corpus-Driven Exploration of MWE
57
thus based on automatically pre-structured data, but is itself a strategy of controlled
human interpretation.
As a main example in this paper, we will look at the MWE Musik in den Ohren,
literally translated as music in the ears, which is similar to the English MWE music
to the ears. Musik is a significant collocation partner of Ohren and the KWIC lines
containing these two word forms in an appropriate distance are grouped in a collocation
cluster. This cluster is the starting point for our analysis. Table 1 shows an excerpt of
the relevant KWIC lines.
M06
MLD
S94
O94
O94
O94
O94
O95
O95
O95
O95
O96
O96
O96
O97
O97
O97
O97
O97
O99
"klingt wie Musik in meinen
"Das ist Musik für unsere
Die Musik kann für westliche
Chancen.. ." Musik in unseren
Uhren,
wahre Musik in Horst Fendrichs
das war Musik in den
Motoren wieder Musik in seinen
Sonderling, Künstler, der "die
sind auch Musik in den
Lust sind Musik in den
ist wahre Musik in seinen
Gästen. Musik in Carsten Kelms
narrative Musik halten Augen und
- zeitgenössische Musik für junge
Während es etwa in den
Musik in des versöhnten Ombudsmans
euch" wie Musik in den
angekündigt. Musik in meinen
der lauten Musik kurz die
Ohren",
Ohren",
Ohren
Ohren.
Ohren,
Ohren.
Ohren
Ohren.
Ohren
Ohren
Ohren
Ohren:
Ohren,
Ohren
Ohren
Ohren
Ohren!
Ohren.
Ohren.
Ohren
sagte die Sozialbürgermeisterin und
sagt Ursula Schmitz. Musik, von
eine Qual sein; sie ist
Doch auch wenn uns wohltut,
Musik In der Grazer ESC
"Es leben die PS.. ."
des Akustik-Sachverständigen. Auch der
Die ersten vier
voll Musik hat und den
des Liebespartners. Was gehört noch
des Liebespartners
Seine Liebe zur Eisenbahn ließ
denn wenn‘s bei denen im
immer wieder lustvoll auf Trab.
(Ossiach, 13. bis 16. 7.).
vieler wie Musik klingt, wenn
Die Starparade ist der Höhepunkt
Aber die Papst-Visite wurde ebenso
zuhielt... Gefeiert wird in St.
Table 1. KWIC lines from the collocation cluster Ohren – Musik
2.2
The level of surface patterns
On this level the KWIC lines which have been grouped by collocation analysis are
subjected to further structuring. For this we use a query syntax based on regular expressions. The queries are used to identify and group lines with a similar syntagmatic
structure which serve then as a basis for the analytic steps that follow.
This step is necessary, as our definition of MWE calls for a common syntagmatic
structure of the instances of an MWE while collocation analysis looks at word form surfaces without considering the syntactic connection between them. Thus, it sometimes
sorts instances of different MWE which share the same lexical material into a single
cluster or assigns instances of the same MWE which have different lexical material
(e.g. because of orthographical variance or different realizations of the same lemma)
to different clusters. Humans, as opposed to the computer, can decide which surface
similarities are important for the task of identifying and describing MWE and formulate
surface patterns designed to gather the correct instances.
Pattern matching is also a valuable asset when exploring the variability of an MWE.
The patterns can be formulated more or less restrictively and the researcher can observe
how many KWIC lines – i. e. instances of realization of the MWE in the corpus –
58
Annelen Brunner et al.
are captured. It is also possible to define gaps in the patterns and study the fillers for
these gaps. This gives a very good indication what features are really important for the
structure of a MWE, how it can be modified and what is its core meaning.
The definition of surface patterns is an iterative process. Often the form of the
patterns has to be adapted when observing the results of the previous try. This process
reflects an ever deepening understanding of what is relevant in the MWE structure.
The cluster Ohren – Musik in Table 1 above is a good example for a collocation cluster which contains instances of different MWE, for example westliche Ohren [western
ears] or die Ohren zuhalten [to cover the ears]. Here are some examples of search
patterns which are used to filter out instances which are relevant for the description of
the MWE Musik in den Ohren. (#* stands for an arbitrary number of unspecified word
forms; N(den) stands for “not den”; ist|war stands for “ist or war”.)
(1)
(2)
(3)
(4)
Musik in #* Ohren
Musik in den Ohren
Musik in #* N(den) Ohren
Das ist|war Musik in #* Ohren
You can see that the search patterns differ in their restrictiveness and that, though for
sake of simplicity the ‘name’ of the MWE was given above as Musik in den Ohren, not
all realizations actually take exactly this form.
Search pattern 3 is specifically designed to capture all realizations that do not use
the definite determiner den. We can now examine the hits of this search pattern and learn
from the filler list for the gap that possessive pronouns and genitive forms referring to
persons are also common when this MWE is realized. However, frequency counts show
that the realization with den appears almost three times as often in our corpus.
Search pattern 4 is an example of a very specific pattern. It captures a common way
of using the MWE which is very stable, though it covers only a relatively small section
of all instances of the MWE in our corpus.
2.3
The level of Lexical realizations
The level of Lexical realizations (LR) is a step between the surface patterns and the
actual MWE. LR represent typical realizations of an MWE in the corpus. This intermediate level of analysis has been introduced to account for the fact that MWE are
very flexible and subject to much variation. When generalizing immediately to a single
typical form, many of these nuances would be lost. LR focus on different kinds of
realization of the same MWE and offer a chance to comment on them.
From empirical experience, we have defined several types of LR. This typology
is quite general and reflects very basic mechanisms of language. The types will be
presented below and exemplified by the LR of the MWE Musik in den Ohren. Figure 2
illustrates how the KWIC lines from the collocation cluster are assigned to several LR
(each subsuming the hits of one or more search patterns) and how the LR tree for this
MWE is built up.
Core LR We assume that there is a core structure which is necessary for the MWE to
be recognizable. This structure is captured in the Core LR. Often, the structure of the
Corpus-Driven Exploration of MWE
59
Fig. 2. LR structure of the MWE Musik in den Ohren in relation to the KWIC lines of the cluster
Ohren – Musik
60
Annelen Brunner et al.
Core LR is also the most general and subsumes the largest number of KWIC lines, i.e.
instances of the MWE in the corpus.
Example: The Core LR Musik in den Ohren subsumes more than half of all the
instances of the MWE and can thus be regarded as the most common realization –
especially as the surface forms of alternative realizations are not as stable.
Core Variant LR The core can be subject to variations on the surface, which are
documented in the Core Variant LR. The Core Variant LR is defined relative to the
Core LR and differs from its structure in at least one respect.
Example: The Core Variant LR Musik in N(den) Ohren subsumes all cases in which
the determiner den is not present. The fillers for the search pattern gap N(den) are
presented prominently in form of a filler list and it becomes clear that especially possessive pronouns or genitive forms referring to persons take the place of the determiner.
It is justified to differentiate between a Core LR and a Core Variant LR instead of just
defining a more general Core LR with the search pattern Musik in #* Ohren, because
it allows us to highlight the fact that these two different types of realization exist and to
show their relative frequencies and nuances in meaning.
Extension LR The structure of the Core LR can be extended by additional elements,
which are not mandatory for the structure of the MWE, but are still frequent and typical
for the way the MWE is used. There can be internal extensions, which appear between
the elements of the core structure, or external extensions, which are added before or
after the core. Extensions have to be connected to the core syntactically. They are for
example verbal constructions, modifiers, object extensions or prepositional phrases.
Example: There are several Extension LR in the LR tree for the MWE Musik in den
Ohren. All of them extend the structure captured by the Core LR and Core Variant LR
in different ways. As you can see in Figure 2, some of them demonstrate that an LR
can subsume more than one search pattern. Usually search patterns that only account
for word order or grammatical variances are bundled together, unless one of them is
extremely prominent or idiosyncratic.
On the first level, there are the LR wie Musik in X Ohren and the LR Musik in X
Ohren sein. Those LR have in common that they add one element to the core structure.
Note that they both capture instances of the Core LR structure as well as of the Core
Variant LR structure – the difference between those structures has been highlighted
already, so it is not necessary to separate them in the Extension LR. The variable
component is marked in the name of the LR by the letter X.
Both Extension LR have subordinate Extension LR which further differentiate the
structure. LR wie Musik in X Ohren can be extended to wie Musik in X Ohren klingen.
This LR is interesting because it is in fact a combination of two MWE: Musik in den
Ohren and in den Ohren klingen [resound in the ears]. Musik in X Ohren sein can be
extended to Das war/ist Musik in X Ohren and wie Musik in X Ohren sein. In fact,
the latter LR could also be defined as a child element of wie Music in X Ohren as it
represents a combination of the two extensions.
Corpus-Driven Exploration of MWE
61
Context LR These LR serve as a focus on typical contexts in which the MWE is
used. They highlight word forms that commonly appear close to the MWE, but are not
directly connected to its structure (as opposed to the elements of the Extension LR).
Context LR show the associative frame of the MWE and are thus useful to understand
the pragmatics of its usage.
Example: The MWE Musik in den Ohren does not have Context LR. However, a
typical Context LR would be Töne ... das menschliche Ohr which belongs to the MWE
das menschliche Ohr [the human ear] and subsumes the search patterns Töne #* das
menschliche Ohr and das menschliche Ohr #* Töne. The Context LR highlights
a word form, Töne [sounds], which appears significantly frequently in the vicinity of
the MWE’s Core LR. This is an indicator that the MWE typically refers specifically to
the human ability of hearing as opposed to other characteristics of the human ear.
LR tree and LR group The different types of LR can be arranged in a hierarchical
structure and may be assigned a narrative comment explaining their specifics. Together,
they give a differentiated picture of the MWE in its realizations according to the observed corpus evidence.
Apart from the specialized LR types listed above, there is also the LR group. It
serves as a container for collecting all relevant instances of an MWE at once and is
always used as the trunk of an LR tree. It captures the overall frequency of the MWE and
also preserves instances of realization which are not frequent enough to be highlighted
by specialized LR, but may still be of interest for a researcher working with our results.
2.4
The level of MWE
An MWE subsumes an LR tree and is assigned a paraphrase that describes a generalized
meaning. The special nature of MWE in our approach becomes evident at this point:
An MWE is not a static form, but a complex set of realizations from which a common
communicative meaning emerges.
As our approach is based on collocation profiles of target word forms, we established the rule that an MWE must contain at least two concrete and immutable lexical
elements – the target word form and one collocation partner. This may seem a somewhat
artificial restriction, but it helps greatly in structuring the rich and often overlapping
structures that can be detected when analysing language in this manner.
To account for the complex interrelations between MWE, it is possible to define
links between them. MWE which have a similar or opposing meaning are connected
and their relationships are commented on. Also, structural overlap between MWE is
pointed out. Links can be defined between MWE which arise from the same collocation
profile (and contain the same target word form), but also between MWE from different
profiles.
The example MWE Musik in den Ohren is assigned the LR tree shown in Figure 2
and a paraphrase if its general meaning is added: “This MWE is used to express that
something is received positively and considered pleasing or beneficial.”
Also, links to related MWE are defined. Within the same profile these are for example the MWE in den Ohren klingen [to resound in the ears]. As mentioned above, these
62
Annelen Brunner et al.
two MWE are commonly combined. There is also a link to the MWE Musik für X Ohren
[music for X ears] which has a quite similar meaning, but is more often used to refer
to actual musical preferences. In addition to that, there is a connection to a different
profile, that of the singular form Ohr [ear] where the MWE Musik in X Ohr [music in
X ear] can be found, which is nearly identical in meaning, but much less frequent.
2.5
The level of MWE patterns
MWE patterns are abstractions over several MWE. This level accounts for the fact that
structures in language can be much more general than the restrictions we imposed on
MWE allow for.
Much regard is given in our model to mutable elements in fixed structures, which
also appear on the level of surface patterns and LR. On the level of MWE patterns,
this concept is brought to a higher degree of abstraction: MWE patterns generalize
over MWE which are structurally similar, but different in some aspect of their lexical
structure. Therefore these patterns always contain underspecified components.
Two types of MWE patterns can be distinguished. In the first case, the realizations
of the underspecified components can be regarded as synonyms. The MWE pattern
structurally generalizes over MWE which carry essentially the same meaning.
In the second case, the realizations of the underspecified components are a set of
dissimilar lexical items. Each MWE subsumed by the MWE pattern has a different
communicative function in its own right, but they share a common core meaning which
can be assigned to the more generalized structure.
MWE patterns are especially interesting from the point of view of construction research as they illustrate the transition from concrete lexical items to abstract structures.
Fig. 3. Excerpt from the hierarchy of the MWE pattern aus ADJECTIVE Gründen
Corpus-Driven Exploration of MWE
63
An example for an MWE pattern of the second type is the structure aus
ADJECTIVEDOMAIN Gründen [for ADJECTIVEDOMAIN reasons]. This MWE pattern
subsumes several MWE where the underspecified component is realized by a specific
lexical item and which each have a different meaning. However, a communicative function which is shared by all its child MWE can be attributed to the MWE pattern: “Using
this pattern makes the actions that are explained seem official and at the same time
allows the speaker to be vague about the reasons for these actions by using the less
specific plural form (Gründen [reasons]) which is mandatory for its structure.”
Of course, aus ADJECTIVEDOMAIN Gründen is itself a specialization of the even
more general MWE Pattern aus ADJECTIVE Gründen. The meaning assigned to this
most abstract MWE pattern is justification in a general sense. Figure 3 shows part of
the hierarchical structure of MWE patterns and MWE.
3
Presentation and prospects
In our ‘Wortverbindungsfeldern’ (MWE fields) [Steyer/Brunner 2008] we have created
a hypertext view which reflects the steps of interpretation detailed in this paper and
gives access to the corpus data the analysis is based on. In this way, it is possible to
completely retrace our interpretative steps and decisions in generalization. The presentation is enriched by comments on the meaning and specifics of MWE patterns, MWE
and LR.
At the moment, the focus of our research is on the development of the model of
analysis. We are planning to extend the network of interrelations especially between
MWE that originate from different collocation profiles and expand the level of MWE
patterns. Also, annotations for MWE should be introduced, which would allow dynamic
grouping and different views.
Another research question is how to adapt the in-depth method for a larger scale
description of MWE and to develop modes of presentation suitable for different user
groups – for example learners of German versus researchers interested in constructions.
References
[Belica 1995] Belica, Cyril (1995): Statistische Kollokationsanalyse und Clustering.
Korpuslinguistische Analysemethode. Mannheim. Internet:
http://www.ids-mannheim.de/kl/projekte/methoden/ur.html1
[Brunner/Steyer 2007] Brunner, Annelen / Steyer, Kathrin (2007): Corpus-driven study
of multi-word expressions based on collocations from a very large corpus. In: Proceedings of the 4th Corpus Linguistics Conference, Birmingham. Internet:
http://corpus.bham.ac.uk/corplingproceedings07/paper/182_Paper.pdf1
[CPA] Corpus Pattern analysis. Internet:
http://nlp.fi.muni.cz/projekty/cpa/1
[Hanks 2008] Hanks, Patrick (2008): Lexical Patterns. From Hornby to Hunston and
Beyond. In: Bernal, Elisenda / DeCesaris, Janet (eds.): Proceedings of the XIII
Euralex International Congress, Barcelona, 15-19 July 2008. Barcelona, p. 89-129.
1
Web sites were checked on 29 July 2009.
64
Annelen Brunner et al.
[Feilke 2004] Feilke, Helmuth (2004): Kontext – Zeichen – Kompetenz. Wortverbindungen unter sprachtheoretischem Aspekt. In: Steyer, Kathrin (ed.):
Wortverbindungen – mehr oder weniger fest. Jahrbuch des Instituts für Deutsche
Sprache 2003. Berlin/New York.
[KLa] Projektseite Ausbau und Pflege der Korpora geschriebener Gegenwartssprache:
Das Deutsche Referenzkorpus – DeReKo. Internet:
http://www.ids-mannheim.de/kl/projekte/korpora/1
[KLb] Projektseite Methoden der Korpusanalyse- und erschließung. Internet:
http://www.ids-mannheim.de/kl/projekte/methoden/1
[Sinclair 1991] Sinclair, John (1991): Corpus, Concordance, Collocation. Oxford.
[Steyer 2000] Steyer, Kathrin (2000): Usuelle Wortverbindungen des Deutschen. Linguistisches Konzept und lexikografische Möglichkeiten. In: Deutsche Sprache 28,
2. Berlin, p. 101–125.
[Steyer 2004] Steyer, Kathrin (2004): Kookkurrenz. Korpusmethodik, linguistisches
Modell, lexikografische Perspektiven. In: Steyer, Kathrin (ed.): Wortverbindungen – mehr oder weniger fest. Jahrbuch des Instituts für Deutsche Sprache 2003.
Berlin/New York, p. 87–116.
[Steyer/Brunner 2008] Steyer, Kathrin / Brunner, Annelen (2008): Wortverbindungsfelder. Abrufbar unter “Wortverbindungen online”. Internet:
http://www.ids-mannheim.de/ll/uwv/wvonline/wvfelder1
[Steyer/Brunner 2009] Steyer, Kathrin / Brunner, Annelen (2009): Das UWVAnalysemodell. Eine korpusgesteuerte Methode zur linguistischen Systematisierung von Wortverbindungen. (= OPAL – Online publizierte Arbeiten zur Linguistik
1/2009). Mannheim. Internet:
http://www.ids-mannheim.de/pub/laufend/opal/privat/opal09-1.html1
[Steyer/Lauer 2007] Steyer, Kathrin / Lauer, Meike (2007): “Corpus-Driven”: Linguistische Interpretation von Kookkurrenzbeziehungen. In: Kämper, Heidrun /
Eichinger, Ludwig M. (eds.): Sprach-Perspektiven. Germanistische Linguistik und
das Institut für Deutsche Sprache. (= Studien zur deutschen Sprache 40). Tübingen,
p. 493–509.
[Tognini-Bonelli 2001] Tognini-Bonelli, Elena (2001): Corpus Linguistics at Work. (=
Studies in Corpus Linguistics 6). Amsterdam/Philadelphia.
Text-Oriented Thesaurus Retrieval System
for Linguistics
Natalia P. Darchuk and Viktor M. Sorokin
Kyiv Taras Shevchenko University, Ukraine
Abstract. The aim of the project is 1) a compilation of the electronic Dictionary
of linguistic terms using a new formalized thesaurus compilation methodology
that meets the standards of the terminography and its presentation in Internet; 2)
a verification of the theoretical thesaurus model owing to the developed
computer technologies by using it for the analysis of texts taken from various
spheres of linguistics.
The importance of the project is that IRS in the multimedia space first of all
provides linguists with a modern standardized dictionary of linguistic terms; secondly,
the result of the project is a methodology of thesaurus compilation and also computer
tools for realization of the methodology; thirdly, IRS is compatible with intellectual
systems of text information processing where it can be implemented as a knowledge
base and a tool for text meaning recognition.
Logic-notional modeling of terminology systems in different sciences and
branches of knowledge is one of the topical inter-branch problems nowadays as the
models of terminology systems are required for compiling terminological dictionaries,
information thesauruses, classifiers, rubricators, creating automated retrieval systems,
databases, artificial intelligent systems. Development of the information-retrieval
thesaurus (IRT) that, on the one hand, is a mean for formalized representation of
terminology, since it represents semantic relations between terms rather strictly, and,
on the other hand, is an important tool for constant improving of knowledge systems
of specific sciences, is considered to be a particular case of knowledge modeling. The
project that is developed in the laboratory of the computational linguistics of the
Philology Institute in the Kiev National Shevchenko University is dedicated to this
problem.
The aim of the project is
• a compilation of the electronic Dictionary of linguistic terms using a new
formalized thesaurus compilation methodology that meets the standards of
the terminography and its presentation in Internet;
• a verification of the theoretical thesaurus model owing to the developed
computer technologies by using it for the analysis of texts taken from various
spheres of linguistics.
66
Natalia P. Darchuk and Viktor M. Sorokin
The project was divided into two phases. At the first stage, we developed
information retrieval system represented as a lexicographic and encyclopedic
electronic database of the Ukrainian linguistic terms that consists of four dictionaries:
alphabetic, translation, explanatory and thesaurus. In the alphabetic dictionary
every term (3400 terms) or terminological word combination has its English
equivalent (it is planned to add Russian, German, French and Italian) and definition
(Figure 1). Thesaurus is a list of logic-semantic relations between linguistic terms
(the list was taken from the work [3] but we supplemented and modified it). So, the
developed IRS includes not only separate terms represented as an alphabetically
arranged list together with their interpretations but also the models that show the
relations between terms (Figure 1).
On the basis of the modern linguistics achievements interpretations of the
terminological units are taken from the reliable sources (about 30 of them) such as
terminological dictionaries, grammars, monographs and given in the compact and
available form. The dictionary includes general linguistic terms – mainly nouns or
noun phrases – from all spheres of grammar, lexicology, applied and computer
linguistics.
Dictionary units are widely used general linguistic terms from morphemic, word
formation, paradigmatic, syntax, lexicology, semantics, terms from particular applied
spheres known in the Ukrainian, Russian and foreign linguistics, terms from computer
linguistics connected with the automation of linguistic processes. Selection of the
dictionary units to the database was based on the heuristic principles (knowledge of
the thesaurus compilers and linguistists-experts).
translation
dictionary
explanatory
dictionary
alphabetically
arranged list
thesaurus
Fig. 1. Fragment of the electronic dictionary of linguistic terms
Text-Oriented Thesaurus Retrieval System for Linguistics
67
The process of the thesaurus compilation is supposed to reveal all types of
relations between notions expressed by terms. The main relations are hyponymy
(genus-species), collateral subordination at one level – partiation (whole - part),
synonymy, correlation, association, object localization, its destination, function,
means of function expression, relations etc. Content of relations is widened so, that it
is possible to cover maximal number of terms connected with the analyzed term as a
dictionary unit. As an interpretation was sometimes not enough to get all essential for
the terms relations, we also took into consideration encyclopedic dictionaries,
scientific works on the particular problem, our own knowledge and knowledge of
linguists-experts. Dictionary entry is structured as a form that was filled for every
term. The form includes a standard list of relations that are notional for a dictionary
unit. The name of a relation is a binary predicate R (A, B) that links the main word of
the entry (A) with the term (B) introduced by this predicate [3, 22].
Thesaurus consists of 3394 terms that are covered by a semantic net with 9265
semantic relations (Table 1).
Type of semantic relations
Term A
(example)
Term B
(example)
Genus – Species (A is generic
to B)
Discipline (А розглядається в
дисципліні В)
(A is regarded in discipline of B)
Synonyms (А is synonymous
to В)
Whole - Part (В consists of А)
Correlate ( А is opposite to В)
Watch…(about А watch В)
частини мови
несамостійні
частини мови
синтаксис
Association (А is associated
with В)
Operation/Procedure (B is an
operation/procedure for A)
Initial object (В is performed
over А)
Tool/Method (А using В)
автоматичне
анотування
аломорф
Parameter (А is characterized
by В)
абзац
автоматични
й переклад
абзац
агент
акузатив
Frequency of
semantic
relations’
realization in the
thesaurus
1843
1157
1016
субморф
машинний
переклад
писемний текст
пацієнс
знахідний
відмінок
морф
абревіатура
абревіація
263
речення
автоматична
сегментація
тексту
мережеве
моделювання
додаткова
дистрибуція
232
934
822
529
312
221
205
68
Natalia P. Darchuk and Viktor M. Sorokin
Parameter carrier (B is a carrier
of parameter A)
Final object (А is performed
over B)
Is related to…(concerns adjective
terms: А is related to В)
Way of expression (А is
expressed by В)
Language level (А is regarded at
the level specified by B)
Main function (А expresses В)
Relation (В restricts А )
Way of the object representation
(А is presented through В )
Aspect (А is regarded in the
aspect В)
Linguistic object (А is presented
as В)
Implication (if А, then В)
Unit of the level (B is a unit of A)
Class (А is a member of class В)
Object of a science (А is an
object of В)
абревіація
іменник
210
автоматични
й синтез
тексту
акузативний
текст
202
акузатив
189
особове
дієслово
актантна
структура
вигук
головний член
особова форма
дієслова
синтаксичний
рівень
експресивність
безпосередня
синтаксична
залежність
дериваційна
історія слова
184
писемний
текст
семантична
мережа
перехідне
дієслово
словоформа
99
словотвірна
структура
слова
графіка
зв’язний
текст
знахідний
відмінок
морфологічни
й рівень
аломорф
природна
мова
морф
інженерна
лінгвістика
176
157
130
109
96
81
40
36
22
Table 1. Quantitative characteristics of semantic relations
The main paradigmatic semantic relations (genus-species, synonymy, whole-part,
correlation), as it can be seen from the table, cover considerable amount of terms
(almost 70% of all paradigmatic relations). From the point of view of theoretical
semantics the more semantic information a dictionary includes, the better, because the
large number of relations in the thesaurus gives a user a greater opportunity to express
his information needs in the request. As the dictionary entry is a synthesis of
linguistic, translation, explanatory and encyclopedic information, we, in connection
with information approach intended for a user request in the IRS, developed an
interactive system (the example is given below):
Request: «what term is a generic for a term “абстрактний іменник” (abstract
noun) ?»
Answer: «A noun “іменник” (noun) is a generic».
Text-Oriented Thesaurus Retrieval System for Linguistics
69
Request: «What terms and relations is a term “абстрактний іменник”
(abstract noun) related to?»
Answer: «“Граматичний рід” (grammatical gender) – relation: parameter
carrier; “Абстрактне значення” (abstract meaning) – relation: parameter carrier,
implication»
The answers to the request are taken from the thesaurus graph that looks like a
semantic net – hierarchically arranged data structure of nodes (terms) and ridges that
represent different types of thesaurus relations – and is given from the thesaurus as a
text (Figure 2).
Relations “node (n) – servant (s)” where a dictionary unit is a “node” that
subordinates are colored black on the screen, and relations “servant - node” where a
dictionary unit is vice versa subordinated are colored blue.
There are terms in the thesaurus that are covered by a ramified network of
semantic relations: sentence – 121 (n = 24, s = 97); word – 62 (n = 30, s = 32) etc.
«servant»
«node»
Fig. 2. Thesaurus graph
Since every relation has its own mark and number in the electronic format, it
facilitates getting data grouped the way a user wants. Represented thesaurus model
can be regarded as a semantic and information description of linguistic terminology
that is modified as
• alphabetically arranged dictionary with definitions;
• sectional terminological dictionary of such disciplines as morphology,
cognitive science, computer linguistics etc;
• dictionary of synonyms;
• dictionary of genus-species relations;
• automatic information reference book etc
70
Natalia P. Darchuk and Viktor M. Sorokin
Network data representation has not only applied necessity but also enables deeper
understanding of logic relations in this science, more precise modeling of the
analyzed terminological system.
Compiled thesaurus is a static model of logic-notional relations between terms of
linguistic metalanguage. This model can be examined from the point of view of
representing dynamic aspects of the scientific knowledge structure, meaning as a
verification of theoretical thesaurus model using it for the corpus analysis of texts
from different spheres of linguistics. The importance of such research is explained by
the fact that knowledge has a textual representation and becomes known from texts.
Text is a different form of the existence of scientific knowledge. Here knowledge
functions as a text semantics, that is knowledge gets gradually a status of scientific
information [1, 50]. There is no doubt that the most complete knowledge about an
object, its features and properties are given in the genre of scientific monographs such
as textbooks, dissertation, because inter-branch relations are examined in details in the
scientific monographs. However, other types of texts such as report papers, articles,
discussions etc, that represent scientific knowledge in a different way, cannot be
disregarded.
Any encyclopedic model of the scientific knowledge is a derivative from a set of
real texts and a representation of these texts at the level of the semantic model.
Semantic model of the language of science is an invariant of separate partial texts’
micromodels relatively to which texts are its variants. Relatively to the set of terms of
the particular science, a logic-notional system of the knowledge field is a model of the
plane of content of the knowledge field. The logic-notional structure of the text
represents the main elements of a semantic paradigmatics of texts. Modeling of
scientific texts’ semantics supposes revealing all types of semantic relations among all
words – notions’ names, their features and properties etc.
Two approaches that give different theoretical and practical results are applied
while investigating a term in the text: from the text to the term (that is terminological
text analysis) and from the term to the text (textual term analysis). Terms are formed
in the text but registered in the terminological system (dictionary). And if in the
functioning sphere the term exists as paradigmatic and syntagmatic variants then in
the terminological system it exists only as paradigmatic variants [2, 152].
On developing methods according to which created thesaurus model of linguistic
terms represented as a hierarchical classification scheme-graph is imposed upon the
vocabulary of a scientific text, a combination of these approaches is supposed to be
promising. As a result we also get a hierarchical classification graph of the specific
analyzed text with absolute frequency of a term usage in the specific text. This
analysis is provided by the second approach – from the terminological dictionary to
the text. Next methodical step is development of an auxiliary list of words on the basis
of the text together with corresponding frequencies without taking into consideration
those terms that were included to the thesaurus graph. Revising the words of the list it
is possible to form a list of words that can be considered to have all the potential to
Text-Oriented Thesaurus Retrieval System for Linguistics
71
gain a status of terms. This is provided by the approach from the text to the term, and
the frequency of usage and contexts enable accumulating scattered terminological
information to solve different terminological problems, in particular to analyze the
reasonability of adding words-pretenders to the compiled dictionary of terms.
This explains developing of the second stage of this project – creating the
dynamic logic-notional model of each particular text from the corpus with texts of
linguistic research area.
The methodology consists of the following main steps: 1) lemmatization and
arrangement of lemmas according to their parts of speech; 2) calculation for each
lemma (noun and adjective) of its absolute usage frequency; 3) creation of the
thesaurus graph of terms for a specific text with their absolute usage frequencies in
the text applying the thesaurus graph of the term system; 4) resolving homonymy of
terms’ meanings (e.g. grammar 1 (a systematic description of the grammatical facts
of a language, that is a system of morphological units, categories and forms,
syntactic units and categories, derivational units and means of word formation);
grammar 2 (the branch of linguistics that deals with grammatical structure of a
language) and grammar 3 (formal grammar that is a part of the automatic text
processing system)) with the help of a context; 5) compilation of an auxiliary word list
with absolute frequencies that was not included to the thesaurus; 6) search for the
words-pretenders on the role of the terms and their registration together with
illustrative context in the additional terminological dictionary. The work of steps 1-3
and 5 in automatic mode and 4 and 6 in automated mode is provided by the algorithm
and program packages. The performance of the methodology was checked on the
corpus of linguistic scientific articles, because the article as one of the forms of
scientific knowledge existence discusses one of the aspects, shows one of the sides of
the scientific notion or problem providing a wide range of scientific knowledge, its
represented terminological semantics (texts of 72 articles from the collection
“Українське мовознавство”; text length is 82418 word usages).
As the result of automatic applying of the thesaurus represented as an oriented
graph to the general word list of the corpus with the linguistic scientific articles we
got another word list also represented as an oriented graph. Considering previously
compiled thesaurus as an invariant model of the linguistic knowledge, relatively to the
analyzed texts thesaurus is a plane of content model: semantic net represents
hierarchically organized data structure consisting of the nodes (terms) – the notions’
names – and ridges that show the relations between these notions (Figure 2).
Word list of 821 terms covers 16,12% of text (13287 terms of 82418 word usages
in the text cumulatively). At the same time about 24% of linguistic terms included to
the thesaurus were found in the corpus (821 terms of 3340 terms in the thesaurus).
The next step was checking whether the automatic auxiliary word list contains
terms or terminological word phrases that were not found in the general linguistic
thesaurus because of the number of reasons: firstly, there is can be author’s terms
included to the text because of the necessity to show the author’s position about some
72
Natalia P. Darchuk and Viktor M. Sorokin
theoretical problem; secondly, dictionary can never be complete as the process of
science-technological terminology development can never be stopped. Text is a
source where a term is formally and semantically formed. Only in the text it is
possible to retrace “the life” of the term solving the question about reasonability of
including it into the terminological dictionary.
Described methodology enables finding and further examining of the term
functioning in the automated mode. Terms in texts are formed together with ideas,
notions and are followed sometimes with such phrases as “by this term something is
meant…” etc. This can be a reason for compiling a list of words-pretenders on the
role of a term together with their contexts enough for solving the problem of
reasonability, or, in case of giving different from generally accepted interpretation, a
the reason for saving them together with the interpretation in the specially developed
terminological data bank (TDB)
Fig. 3. Thesaurus of terms from the collection of articles
«Українське мовознавство»
The auxiliary word list contains terms from the following branches of linguistics:
Stylistics: фоностилема, стиледиферентор, іностильове включення,
ритмічні синтагми, ритмізація синтагми, горизонтальна / ланцюжкова
ритмізація текстів, вертикальна ритмізація тексту (В.Берковець).
Cognitive science: домовна картина світу, просторова картина світу,
національно-мовна картина світу, міжнародно-правова картина світу,
асоціативний ґештальт (І. Скорик); ядро ґештальту, зони ґештальту
(Д.Терехова); парадигма мислення, когнітивна модель, когнітивна
діалектологія (Н.Руснак), інгерентний аспект конотації, адгерентний аспект
конотації (І.Калита).
Text-Oriented Thesaurus Retrieval System for Linguistics
73
Lexical science категорійна система лексикону школяра (В.Чумак,
А.Акуленко); колоквіалізм (І.Ковалинська); народні компаративеми,
еліптування (О.Дуденко); інтерпретаційна модель оксюморонів (В.Мерінов);
загальний сленг, спеціальний сленг (В.Вітренко); уніноміальні назви вищих
таксонів (С.Руденко).
Semantics: субкоди семантичної субстанції (Л.Скиба), психосемантика
(Т.Печончик).
Syntax: хронотип, компаративема (В.Коломийцева); граматика
поетичних інновацій, експресивний синтаксис (Н.Гуйванюк).
Morphology: міноративні (форми, суфікси), міноративність (Н.Руда);
аглютинант (С.Климович).
Individual author’s terms can also be found sometimes. But formed once
individual terms can become a social phenomenon if the theory, where this term
appears, becomes generally accepted being practically confirmed (e.g. the term
“performative verbs” by John Ostin is used in the papers of the Ukrainian grammarian
I.R. Vuhovanec and in the article written by I. Shevchyk.)
The research showed that terms that were not included into the thesaurus are either
rarely used in linguistics (ґештальт, хронотип, таксон, субфрейм, суперфрейм),
or author’s ones (міжнародно-правова картина світу, домовна картина світу,
колоквіалізм, народні коппаративеми, психосемантика etc), and in any case
they were included with definitions provided by the authors of the scientific articles.
Such ways of word formation can be retraced among one word innovations:
• stem-composition (фоностилема, стиледиферентор, психосемантика,
хронотип): an object is designated according to the set of features;
• usage of Greek-Latin patterns (компаратив-ем-а,колоквіал-ізм, фоностилем-а, аглютин-ант, міфолог-ізм);
• abbreviation (ККС – концептуальна картина світу, ПКС – просторова
картина світу, РС – ритмічні структури, ЕКС – емоційно-смисловий
концепт).
There are more two-component nominative word combinations that consist of a
designated word (usually a noun) and a determinant word (adjective, another noun).
Multicomponent word combinations tend to transform into the abbreviation (МКС –
мовна картина світу). In such combinations a determined component is usually a
name of specie that is differentiated by adding a determinant:
1. картина світу (worldview);
2. міжнародно-правова картина світу (international worldview);
3. концептуальна картина світу (conceptual worldview);
4. національно-мовна картина світу (national language worldview);
5. двомовна картина світу (bilingual worldview);
6. просторова картина світу (spatial worldview).
74
Natalia P. Darchuk and Viktor M. Sorokin
The main term (genus: картина світу (worldview)) and derivative from
terminological element, that determines a feature of the specific notion, forms the
whole terminological paradigms complicating the basic notions that are lately being
used in the cognitive science papers.
Authors used the following methods while term forming:
1. term formation on the basis of the international terminological elements –
morphological components;
2. loanword of terms – usage of transliterated foreign terms. That means that the
same laws are applied both for term formation and for the formation of any other
lexical units of the Ukrainian language. The typical approach is term combination.
Let’s consider the structure of the terminological thesaurus compiled on the basis
of the linguistic articles and some of the results. It should be mentioned that
percentage of specific and non-specific vocabulary in the word list is approximately
21,3% of the non-terms – generally used words and 78,7% of the terms. Such high
percentage of terms evidences that the chosen texts are using and fixating the terms.
Moreover, architectonics of the text should also be taken into consideration.
Architectonics means a typical structure of the article as a textual genre where the
subject, preconditions, reasons of the research, novelty, author’s position, sources of
the research, experiment methodology, experiment description, results of its
discussion, examples, conclusions, results are highlighted. That is why the modeling
of semantics was oriented towards previous logic-notional, conceptual content
analysis of the scientific papers (not only dictionaries but also monographs, textbooks,
articles were used during thesaurus compilation before analysis). The thesaurus
represented as an alphabetical-frequency tree where terms are the nodes and the points
head subtree that consists of the nodes of the higher levels of the hierarchy (total
absolute frequency either of the whole subtree or the frequency of the particular term
in the branch is given in the brackets).
For example,
Аналіз (95) (Analysis)
Аналіз тексту (1) (Text analysis)
КА (CA) (6) (абревіатура: контекстний аналіз) (abbreviation: context analysis)
Концептуальний аналіз (1) (Conceptual analysis)
Логічний аналіз (1) (Logic analysis)
Семантичний аналіз (1) (Semantic analysis)
The term analysis heads the subtree and its absolute frequency (95) includes the
frequencies of other nodes at the lower levels of the hierarchy. So, the frequency of
this term only is 85.
The frequency terminological dictionary (its upper part) served as the subject
model. The most frequent terms represent the subject dominant of texts. They are
мова (language) (624), слово (word) (360) форма (form) (345), значення
Text-Oriented Thesaurus Retrieval System for Linguistics
75
(meaning) (284), семантика (284) текст (249) (text semantics), мовлення
(speech) (213), структура (structure) (205), іменник (noun) (195), лексика
(vocabulary) (183), ознака (feature) (172), прикметник (adjective) (143),
предикат (predicate) (134), речення (sentence) (133), система (system) (168),
суфікс (suffix) (153), функція (function) (163) etc.
The terminological system of the investigated texts includes several groups of
terms. First of all, they are basic terms of methodological sciences, for example, of
philosophy (аналіз (analysis) – 95, відношення (relations) – 80, система (system)
– 168, форма (form) – 345), logic (концепт (concept) – 154, суб‘єкт (subject) – 57,
об‘єкт (object) – 96, адресат (addressee) – 55, фрейм (frame) – 23), computer
science (інформація (information) – 69, функція (function) – 163, конструкція
(structure) – 92, модель (model) – 78). The terminological system includes the main
terms that designate the notions of the linguistic branch (іменник (noun),
прикметник (adjective), дієслово (verb) (71), речення (sentence), лексема
(lexeme) (92), слово (word) (360) etc.), and its derivatives (e.g. основа (basis) (240):
граматична основа (grammatical stem) (1), лексична основа (lexical stem) (1),
основа дієслова (verb stem) (1), основа іменника (noun stem) (1), основа слова
(stem) (2), основа словоформи (wordform basis) (1), складна основа (complex
basis) (2), непохідна основа (underived basis) (1), похідна основа (derived basis)
(19), основоскладання (stem composition) (3)) that form lexical terminological
paradigm represented by a word combination.
So-called general scientific terms don’t differ from the highly specialized terms of
the methodological sciences by their form, but semantics of general scientific terms
differentiates getting additional semes.
Модель (Pattern) (general scientific term) (78)
Мовні моделі (Language patterns) (2)
Моделі мовленнєвої діяльності (Speech activity patterns) (1)
Модель речення (Sentence pattern) (1)
Синтаксична модель (Syntactic pattern) (1)
Словотвірна модель (Word-building pattern) (1)
In the terminological vocabulary of the linguistic articles the number of one word
terms (524) is twice as large as the number of two word terms (269) and three word
terms (22). That can be explained by the fact that the terminological system of
linguistics is an old one as well as linguistics itself. However, the expansion of the
system due to the new word combinations mainly of the adjective + noun type occurs
as the result of the science development and mutual integration of different subjects.
This tendency occurs when there is an absence of clarity while designating some
notions with one word. In this case additional components are being added to it. If
designation has redundant clarity or accuracy then it is being cut forming abbreviation
(e.g. translation – machine translation - MT).
Thesaurus representation of the terminological system as a semantic net is also a
part of the science language description. IRT is a lexical structure pattern as the
76
Natalia P. Darchuk and Viktor M. Sorokin
thesaurus reflects the paradigmatic structure of vocabulary in an explicit form and
enables systematizing all main vocabulary of the investigated material (corpus of the
linguistic articles in this case) from the paradigmatic point of view.
The thesaurus illustrates semantic continuity: dictionary cannot contain
semantically isolated terms. Every term is related to the meanings of the other by the
relations of genus-species, partiation, synonymy and correlation. It is possible to
move from the semantic net to the subject one on the following bases: the more terms
(with their quantitative indexes) in the text are connected in the thesaurus tree by the
semantic relations, the greater subject significance they have in the text. Subjects are
the names of the subject groups in the thesaurus tree. On the other hand, the subject
net is a hierarchy of the main and additional subjects. The following methodological
statements are the initial: 1) thematically significant terminological groups are being
developed on the basis of the thesaurus tree; 2) the hierarchy of subjects of the text is
determined depending on the number of terms included into subject groups.
Addition of a new scientific text in the linguistic corpus causes updating of the
IRT model. This model enables: 1) automatic selection of the terminological lexemes
on the basis of a coincidence with a general thesaurus; 2) automated control provided
by a specialist: identifying terms that are not included into the thesaurus; giving them
definitions, translations, establishing thesaurus connections.
The process of lexico-subject model development consists of the three stages:
1. lemmatization of the text word forms, compilation of the frequency word list of
the text and its arrangement according to the parts of speech;
2. calculation of the absolute frequency that shows the number of terms met in the
branch and the general frequency of the term in the texts;
3. homonymy resolution of the terminological lexemes and enumeration of the
frequency characteristics.
Significance of the corpus research of the terms’ functioning is beyond dispute, as
scientific theories, concepts are outlined and methods are described in the texts. And
terms are then included into dictionaries (or standards, classifiers). So, the main
principle in the science of terminology is the initial sphere of term functioning in socalled term forming and term using texts.
The following methods were used during terminological text analysis and textual
term analysis:
• logic method of classification and construction of the terminological system
establishing relations between terms;
• logic, linguistic, terminological methods applied to identify the facts of combined
usage of terms together with non-terms and general scientific terms, analyze the
combinations of terms with terms in the text;
• quantitative methods applied to identify the importance of the term.
Text-Oriented Thesaurus Retrieval System for Linguistics
77
Compiled text-oriented IRT with linguistic instrument and software can be applied
while developing terminological data banks (TDB) and terminological knowledge
banks (TKB), since a large amount of terminological information has been worked
out (3400 units), tools to automate the processes of selecting, keeping and searching
this information: dictionary of linguistic terms is kept in the computer memory and
can be developed and supplemented by new words. Advantages of such dictionary of
a traditional paper one lie in the fact that electronic IRT enables entering through any
characteristic of a term. On the other hand, the term is not restricted by its dictionary
entry like in the paper dictionary – logical and associative relations are established
between terms in the TDB: relations of genus-species, partiation, synonymy,
correlation etc. Semantic nets, that explicitly represent the structure of the multilevel
terms’ classifications within branch linguistic terminological system, are developed
with the help of the programs.
The importance of the project is that IRS in the multimedia space first of all
provides linguists with a modern standardized dictionary of linguistic terms; secondly,
the result of the project is a methodology of thesaurus compilation and also computer
tools for the methodology realization; thirdly, IRS is compatible with intellectual
systems of text information processing where it can be implemented as a knowledge
base and a tool for text meaning recognition (automatic abstracting and annotating of
the scientific linguistic texts).
References
[1] Герд А. С. Прикладная лингвистика. – Издательство С.-Петербургского университета. – 2005. – 266 с.
[2] Лейчик В. М. Терминоведение: Предмет, методы, структура. – М.:
КомКнига, - 2006. – 256 с.
[3] Никитина С.Е. Тезаурус по теоретической и прикладной лингвистике.
– М., 1978.
From Electronic Corpora to Online Dictionaries
(on the Example of Bulgarian Language Resources)
Ludmila Dimitrova
Institute of Mathematics and Informatics, Bulgarian Academy of Sciences, Bulgaria
Abstract. The paper briefly describes Bulgarian digital language resources,
among them corpora, lexical databases, lexicons, and electronic dictionaries,
which were developed in the Mathematical Linguistics Department at the
Institute of Mathematics and Informatics of the Bulgarian Academy of
Sciences (IMI-BAS) in the framework of some international projects. The
first Bulgarian electronic corpora and language-specific resources were
developed in the EC language technology project MULTEXT-East
Multilingual Text Tools and Corpora for Central and Eastern European
Languages. The first lexical database for Bulgarian was developed in the EC
project CONCEDE Consortium for Central European Dictionary Encoding.
These recourses were developed in TEI-format, and thus they are compatible
with other TEI-conformant resources. The first Bulgarian-Polish electronic
corpora and dictionaries are currently developed in the frame of bilateral
collaboration between IMI-BAS and ISS-PAS.
1 Introduction
The Department of Mathematical Linguistics at the IMI-BAS has successfully
participated in the EC language technology projects MULTEXT-East and CONCEDE.
The MULTEXT-East project (MTE for short: [2]), as a continuation of the project
MULTEXT Multilingual Text Tools and Corpora [10], aims at testing and adaptation
of language standards and corpus tools, developed through the MULTEXT, the
development of language-specific resources for six new languages, and the extension
of the annotated multilingual MULTEXT corpus. MTE developed digital language
resources for six Central and East European (CEE) languages: Bulgarian, Czech,
Estonian, Hungarian, Romanian, and Slovene, as well as for English. These resources
contains, for the each of the six CEE languages, morphosyntactic specifications, lexica,
and corpora. The corpora consist of three parts: parallel, comparable and speechcorpus. Developed in the frame of the MTE project, Bulgarian language digital resources include morphosyntactic specifications, lexicons, and corpora, incl. a parallel
corpus, based on George Orwell's novel 1984, and a comparable corpus [6].
2 Bulgarian language-specific resources: morphosyntactic
specifications
The MTE morphosyntactic specifications have been developed on the basis of the
MULTEXT specifications for Western European languages and in accordance with
the EAGLES guidelines [11]. The morphosyntactic specifications have been used in
the encoding of the word-form lexica of the project. They contain the list of defined
From Electronic Corpora to Online Dictionaries
79
categories – parts of speech (POS), each POS encoded by a letter: noun – N, verb –
V, adjective – A, pronoun – P, determiner – D, article – T, adverb – R, adposition –
S, conjunction – C, numeral – M, interjection – I, residual – X, abbreviation – Y,
particle – Q. A table of attribute-values is defined for each category in order to
reflect the characteristic features of each language. The specific features of each
language are marked up additionally by l.s. The characters following the POSencoding give the values of the position-determined attributes. The specifications
define, for each part of speech, its appropriate attributes and their values, encoded
by one symbol code. It should be noted that if a certain attribute is not appropriate
(1) for a language, (2) for the particular combination of features, or (3) for the word,
this is marked by a hyphen in the attribute's position.
The MTE use the MULTEXT format of lexical description – morphosyntactic
description (MSD), which consists of linear strings of characters, representing the
morphosyntactic information for each word-form. The string is constructed in the
following way:
•
the positions of a string of characters are numbered 0, 1, 2, etc.
•
he agreed character at position 0 encodes the corresponding part of speech:
N for noun, V for verb, etc. ;
•
each character at position 1, 2, n, encodes the value of one attribute (for
nouns the attributes are: type, person, gender, number, etc.);
For example, the MSD Ncfs- means POS: noun, Type: common, Gender:
feminine, Number: singular, nocase.
The proposed formalism for the MSD is not arbitrary (a MSD contains the full
description of a lexical item), but has a clear and concrete aim – to be used for
specific applications, incl. corpus annotation. On the basis of these standard MSDs
the set of corpus tags were determined. A mapping from the morpho-syntactic
information, contained in the lexical description, to a set of corpus tags is also
provided, according to the MULTEXT tagging model. The list of MSDs for
Bulgarian contains 326 elements.
For example, the MSD of the word картата /the map/ is Ncfs-y that means
POS: noun, Type: common, Gender: feminine, Number: singular, nocase,
Definiteness: yes.
Some of MSDs for Bulgarian are not strictly adequate to the particular
morhosyntactic properties of the respective parts-of-speech – Tense, Number,
Gender, Voice, Definiteness – especially in the system of impersonal verbal forms
(participles) [12]. In particular, the present active participial cannot possess the
Tense attribute because it expresses the property/attribute independently and
regardless of the tense of the main verb in the sentence, whereas the Voice attribute
is also implicit from the context. New MSDs for Bulgarian participles have been
proposed, bringing the morphosyntactic description in line with the grammatical
characteristics of the Bulgarian, [7]. An update of the MSDs will make them more
useful for annotation of corpora and automatic disambiguation of Bulgarian texts.
80
Ludmila Dimitrova
The Bulgarian language-specific resources also include a set of segmentation
and morphological rules and data, which are necessary for use with the various
annotation MULTEXT tools. Segmentation rules describe the form of sentence
boundaries, quotations, numbers, punctuation, capitalization, etc. Morphological
rules, needed by the morphological tools, provide exhaustive treatment of inflection
and minimal derivation. The so called special tokens, required by the segmenter,
includes lists of special tokens (frequent abbreviations and names, titles, patterns for
proper names, etc.) with their types. Since some subtools, for example, in the
segmenter require certain language-specific information in order to accomplish their
tasks, each participating side has developed a set of resource files for their
language. For maximum flexibility and to retain language-independence, all such
information is provided directly to the subtools via external resource files.
3 Corpora
MTE is building an annotated multilingual corpus, composed of three major parts:
Parallel Corpus, Comparable Corpus, and a small Speech Corpus of spoken texts
in each of the six languages, comprising forty short passages of five thematically
connected sentences, each spoken by several native speakers, with phonemic and
orthographic transcriptions. The multilingual parallel corpus, based on George
Orwell’s novel “1984” in the English original and the six translations in Bulgarian,
Czech, Estonian, Hungarian, Romanian and Slovene of the novel, was developed.
The parallel corpus is produced as a well-structured, lemmatized, CES-corpus [9].
The corpus contains four parts, corresponding to the different levels of annotation:
the original text of the novel, the CesDOC-encoding (SGML mark-up of the text up
to the sentence-level), the CesANA-encoding (containing word-level morphosyntactic mark-up), and the aligned versions in CesAlign-encoding (containing links
to the aligned sentences). The texts were automatically annotated for tokenization,
sentence boundaries, and part of speech annotation, using the project tools, and
validated for sentence boundaries and alignment. The alignment between the
English version and a translation in each of the six CEE languages ensures six pairwise alignments.
The next examples show excerpts of the Bulgarian-English aligned texts –
Bulgarian-English Aligned 1984 Sampler:
1-1 Aligned sentences:
•
•
<Obg.1.1.7.4>Още три сгради, подобни по външен вид и размери, бяха
посети из Лондон.
<Oen.1.1.9.2>Scattered about London there were just three other buildings of
similar appearance and size.
•
<Obg.1.1.7.5>И дотолкова се извисяваха над околните здания, че от
покрива на жилищен дом Победа можеха да се видят и четирите
едновременно.
•
<Oen.1.1.9.3>So completely did they dwarf the surrounding architecture that from
the roof of Victory Mansions you could see all four of them simultaneously.
From Electronic Corpora to Online Dictionaries
81
1-2 Aligned sentences:
•
•
<Obg.1.1.7.3>От мястото си Уинстън можеше да прочете изписани с
елегантни букви върхубялата фасада трите лозунга на партията:
"Войната е мир""Свободата е робство""Невежеството е сила" Говореше
се, че в Министерството на истината има три хиляди стаи над земята и
съответните лабиринти отдолу.
<Oen.1.1.7.3>From where Winston stood it was just possible to read, picked out
on its white face in elegant lettering, the three slogans of the Party: "War is
peace""Freedom is slavery""Ignorance is strength."<Oen.1.1.9.1> Ministry of
Truth, contained, it was said, three thousand rooms above ground level, and
corresponding ramifications below.
Bulgarian MTE parallel corpus
The Bulgarian parallel corpus contains the Bulgarian translation of Orwell’s
novel “Nineteen Eighty-Four”, includes 86020 words (lexical items, excluding
punctuation), 101173 tokens (words and punctuations); the CesDOC-encoding of
the Bulgarian text of the novel (SGML mark-up of the text up to the sentence-level)
includes 1322 paragraphs, 6682 sentences; the CesANA-encoding of the Bulgarian
text of the novel (containing word-level morpho-syntactic mark-up), and the
Bulgarian-English aligned texts – the aligned versions in CesAlign-encoding,
containing links to the aligned sentences (see examples bellow). The CesANAencoding for Bulgarian in addition includes disambiguated lexical information for
the 86020 words of the novel and undisambiguated lexical information for 156002
words. What is more – there are 156002 occurrences of MSDs in the text (Bulgarian
MSD are 326) and 242 022 occurrences of base or lemma of tokens (which is the
total of 86020 words and 156 002 occurrences of MSD). The number of occurrences
of ctags is 257 175. Each word-form is associated with the respective grammatical
information and the corresponding lemma which form its standard lexical
description.
An example of the CesANA-encoding of the Bulgarian text of “1984” follows:
word-form “и” /and – conjunction/, /so! oh! - interjection/
<tok type=WORD from='Obg.1.1.1.1\24'>
<orth> и </orth>
<disamb><base> и </base><ctag>CC</ctag></disamb>
<lex><base> и </base><msd>Ccs</msd><ctag>CC</ctag></lex>
<lex><base> и </base><msd>I-s</msd><ctag>I</ctag></lex>
</tok>
Bulgarian-Polish parallel corpus
The first Bulgarian–Polish corpus (currently under development) is a result of
the joint collaborative project “Semantics and contrastive linguistics with a focus on
a bilingual electronic dictionary” between IMI—Bulgarian AS and ISS—Polish AS,
coordinated by L. Dimitrova and V. Koseska. It contains a total of approximately
5 million words and comprises two corpora: parallel and comparable [3]. The first
82
Ludmila Dimitrova
Bulgarian–Polish parallel corpus contains more than 3 million words mainly works
of Bulgarian and Polish authors – short stories, novels, children’s literature, science
fiction. A small part comprises official documents of the European Commission
available through the Internet. The corpus is composed of two parts: original
Bulgarian texts with Polish translations or vice versa and texts in other languages
translated into both Bulgarian and Polish.
The corpus is developed according to the MTE model. Most texts have been
annotated at paragraph level. The corpus provides samples for the experimental
version of the Bulgarian–Polish digital dictionary.
In the framework of the joint collaborative project „Electronic corpora –
contrastive study with focus on design of Bulgarian-Slovak digital language
resources“ between IMI—BAS and ĽŠIL—Slovak AS, coordinated by L. Dimitrova
and R. Garabík a small Slovak-Bulgarian parallel corpus is currently under
development.
A small parallel corpus with Bulgarian, Polish, Slovak, Slovene (incl. English as
a hub language) texts of official documents of the European Commission available
through the Internet is also currently collected.
Bulgarian comparable corpora
Bulgarian MTE comparable corpus: For each of the six MTE CEE languages, a
comparable corpus was developed. It included two subsets of at least 100,000 words
each, consisting of
•
fiction, comprising a single novel or excerpts from several novels;
•
newspapers.
The data was comparable across the six languages, only in terms of the number
and size of texts. The entire MTE multilingual comparable corpus was prepared in
CES format, manually or using ad-hoc tools. The Bulgarian comparative corpus
includes Fiction (texts from contemporary Bulgarian literature) and Newspapers
(newspaper excerpts) subsets. The Bulgarian Fiction and Newspapers subsets were
annotated manually. The data in the table below have been determined on a base of
the Bulgarian fiction and Bulgarian newspapers lexica:
Part
word
occurrences
distinct
words
distinct MSDs
in text
distinct Ctags
in text
Fiction
97251
17061
313
129
Newspapers
96538
20696
295
126
The first Bulgarian electronic corpus is included in the MTE multilingual corpus
of the MTE project (http://nl.ijs.si/ME), distributed on CD-ROM by
Trans-European Language Resources Infrastructure (TELRI) Concerted Action
Copernicus 1202, (http://www.ids-mannheim.de/telri/) for research
purposes.
From Electronic Corpora to Online Dictionaries
83
Bulgarian comparable corpus in Bulgarian-Polish corpus: This corpus contains
approximately two million words from works of Bulgarian authors, including prose:
Dimitar Talev, Dimitar Dimov, Pavel Vezhinov, Yordan Radichkov, non-fiction:
Zhelyu Zhelev’s „Fascism“, Bulgarian translations of novels and short stories of
prominent European authors.
4 Bulgarian lexical databases
CONCEDE Bulgarian LDB
The first lexical database (LDB) for Bulgarian was developed in the framework of
CONCEDE project. The lexical databases of the project CONCEDE were
developed on the basis of the MTE parallel multilingual corpus (so-called Orwell
corpus). The CONCEDE project suggested a model for dictionary encoding
containing a lexical database with standardized and well-understood structure and
semantics. The CONCEDE project has developed lexical databases (LDBs) in a
general-purpose document-interchange format for the same six МТЕ CEE
languages: 3000-headword lexical databases for Bulgarian, Czech, Estonian,
Hungarian, Romanian, and a 500-word one from the English-Slovene dictionary.
The project has produced lexical resources that respect the guidelines of the Text
Encoding Initiative - Dictionary Working Group (TEI-DWG), and so are compatible
with other TEI-conformant resources.
The initial word lists for selection of headwords and word frequency were
obtained from the MTE parallel corpus. The selection of headwords was made after
word frequency and word class (POS) were taken into account, and the number of
words there were in a given word-class and word-frequency band.
In order to achieve a harmonization of the LDBs according to the principal
breakdown of lemmata to POS, the CONCEDE consortium decided on the
following proportion: open parts of speech (nouns, verbs, adjectives, adverbs) – no
more than 90 %, closed parts of speech (numerals, pronouns, conjunctions, prepositions, particles and interjections) – minimum 10% of the whole set of lemmata
chosen. Under the CONCEDE project was developed an encoding scheme for
lexicographic specifications of the Bulgarian language, according to the standards
for electronic dictionary encoding [5]. This encoding scheme served to create the
Bulgarian dictionary in the LDBs of CONCEDE. The choice of dictionary entries
follows the method accepted by CONCEDE. The entries are equipped with
lexicographic specifications for Bulgarian language in TEI-conformant SGML. The
electronic dictionary is based on the Bulgarian Explanatory Dictionary [1]. Each
entry is represented as a tree-structure. The chosen entries are divided in the
following POS: noun – 33.84% of the Bulgarian sample; verb – 21.99%; adjective –
12.52%; adverb – 11.51% -- total open POS 79.86%; and numeral – 1.52%;
pronoun – 5.24%; conjunction – 4.06%; preposition – 3.55%; particle – 4.40%;
interjection – 1.35% -- total closed POS 20.13%. The entries in Bulgarian LDBs
retain as much as possible the structure of the original paper dictionary.
84
Ludmila Dimitrova
The example shows the entry in the printed Bulgarian Explanatory Dictionary
with the headword “име” //name//:
име ср. Отличително название на човек, животно и др. прен.
Известност. Той има голямо име. грам. Категория думи, които означават
предмети, качества, числа. Съществително име. Прилагателно име.
Числително име. ◊ В името на предл. Въз основа на, заради. В името на
закона. В името на свободата.
The corresponding entry in the Bulgarian LDBs:
<entry><hw>|име</hw>
<gen>ср.</gen>
<struc type="Sense" n="1">
<def>Отличително название на човек, животно и др.</def></struc>
<struc type="Sense" n="2"><usg type="register">прен.</usg>
<def>Известност.</def>
<eg><q>Той има голямо име.</q></eg></struc>
<struc type="Sense" n="3"><usg type="register">грам.</usg>
<def>Категория думи, които означават предмети, качества, числа.</def>
<eg><q>Съществително име.</q></eg>
<eg><q>Прилагателно име.</q></eg>
<eg><q>Числително име.</q></eg></struc>
<struc type="Phrases">
<struc type="Phrase" n="1"><orth>В името на</orth><pos>предл.</pos>
<def>Въз основа на, заради.</def>
<eg><q>В името на закона.</q></eg>
<eg><q>В името на свободата.</q></eg></struc></struc>
</entry>
In the final phase of the development of the CONCEDE LDBs an examination
was carried out – a validation process which takes two forms, “formal validation”
and “content validation”. The formal validation was a matter of ensuring that the
databases were valid SGML documents and for the Bulgarian LDBs has been done
by means of a validating SGML-parser. The content validation of the entries
required human intervention and therefore was performed manually.
LDB supporting Bulgarian-Polish online dictionary
The formal model of the LDB [4] supporting the first Bulgarian-Polish
dictionary is the CONCEDE model for dictionary encoding, [8]. The hierarchical
structure of the dictionary entry is a tree-structure and described by 3 structural tags:
entry, struc, and alt. The content tagset includes tags, fully describing the entry’s
content: the grammatical information about the headword, the translation equivalence in Polish, examples of the word’s usage with translation, phrasal usage with
translation (if possible) or explanation, the word’s etymology (if known).
From Electronic Corpora to Online Dictionaries
85
For a more adequate description of the Bulgarian verbs, two new tags are being
introduced to represent the verb’s conjugation (Bulgarian verbs are divided into 3
conjugations): conjugation – a new tag is added to represent the conjugation of
verbs; its structure allows the subtag type for the possible types of conjugations of
Bulgarian verbs. Furthermore, it is allowed to input additional information in the
gram tag for the aspect – perfect and progressive (imperfect) of verbs, and in subc
tag – for transitivity/intransitivity of verbs.
The selection of headwords included in the dictionary’s LDB is based on the
Bulgarian-Polish parallel corpus: the main forms (lemmata) of the most frequent
word forms in the corpus are selected. The word distribution according to POS also
follows the CONCEDE model: open parts of speech - no more than 90 %, closed
parts of speech – minimum 10% of the whole set of lemmata chosen.
The representations of three Bulgarian verbal forms as entries in the LDB
follow:
подчертава|м, -ш vi. podkreślać /underline/
подчертан part. podkreślony /underlined/
подчерта|я, -еш vp. v. подчертавам
<entry>
<hw>подчерта’ва|м</hw>
<pos>v</pos>
<gram>i</gram>
<subc>transitive</subc>
<conjugation>
<orth>-ш</orth>
<type>III</type>
</conjugation>
<struc type="Sense" n="1">
<trans>podkreślać</trans>
</struc>
</entry>
<entry>
<hw>подчерта’н </hw>
<pos>part</pos>
<alt>
<pos>adi</pos>
</alt>
<struc type="Sense" n="1">
<trans>podkreślony</trans>
</struc>
</entry>
86
Ludmila Dimitrova
<entry>
<hw>подчерта’|я</hw>
<pos>v</pos>
<gram>p</gram>
<subc>transitive</subc>
<conjugation>
<orth>-еш</orth>
<type>I</type>
</conjugation>
<xr>подчерта’вам</xr>
</entry>
Transformation of the Lexical Database to the Relational Database is carried out
with the help of tables, into which the search data and indices are input. This
organization allows an automatic creation of a dictionary entry for a Polish word,
whenever the translation equivalence is one-to-one. Of course, the input of
information about the Polish word must be done additionally.
Column / Word
подчерта’ва|м
подчерта’н
подчерта’|я
id
668
669
670
подчерта#ва
подчерта#н
подчерта#
подчертан
подчертая
homonym_index
bg_word
suffix
м
bg_word_search
подчертавам
я
plural
is_plural_rare
conjugation
ш
еш
conjugation_type
3
1
has_gender
gender_feminine
gender_neuter
id_explanation
id_bg_word
668
referent_bg_word
подчерта#вам
Table bg_word
From Electronic Corpora to Online Dictionaries
id
id_bg_word
pl_word
sense_ alternative_
index sense_index
1117
668
podkreślać
1
1
1118
669
podkreślony
1
1
latin_
translation
87
id_explanation
Table pl_word
id_bg_word
id_characteristic
668
17
668
57
669
44
670
18
670
57
Table mm_bg_word_characteristic
id abbreviation_ abbreviation_
bg
pl
description_
bg
description_ descrip
id_
pl
tion_ characteristic
lat
_
type
17
мин. нсв.
vi
глагол от
несвършен
вид
5
18
мин. св.
vp
глагол от
свършен вид
5
44
прич
part
причастие
6
57
прех
transitive
преходен
глагол
7
Table characteristic
The LDB of the Bulgarian-Polish dictionary could be used for the design and
creation of new bilingual online dictionaries in the future.
5 Digital dictionaries
Monolingual: Bulgarian MTE lexica
The Bulgarian MTE lexicons (three in total) cover completely the available texts:
George Orwell's novel 1984, newspaper excerpts and texts from contemporary
Bugarian literature, which form Bulgarian MTE comparable corpora. Bulgarian
Orwell’s lexicon is a lexical list, containing 55200 entries among them 17567
lemmata, needed for use in conjunction with the morphological analyser.
The table below represents the number of lemmata and entries, distributed
according to a POS-characteristic, appeared in Orwell's novel 1984:
88
Ludmila Dimitrova
POS
Lemmata
Entries
Nouns (total)
9891
47969
Nouns - masculine
4180
25100
Nouns - feminine
4120
16493
Nouns - neuter
1591
6376
Verbs
4140
226666
Adjectives
2155
19397
Pronouns
92
110
Adverbs
790
790
Adpositions
98
98
Conjunctions
76
76
Numerals
67
67
Interjections
172
172
Particles
86
86
Total
17567
295431
Each element of the lexicon (one entry per line) contains the following
information: the inflected-form (word-form), the corresponding lemma and its
standard lexical description (MSD) and has the following form:
word-form <TAB> lemma <TAB> morphosyntactic description
An excerpt from the Bulgarian lexicon follows:
Word-Form
бели
бели
бели
бели
бели
бели
белите
белите
белия
белият
белота
белота
белота
белотата
Lemma
беля
беля
беля
беля
беля
бял
беля
бял
бял
бял
=
белот
белот
белота
MSD
Ncfp-n
Vmia2s
Vmia3s
Vmip3s
Vmm-2s
A---p-n
Ncfp-y
A---p-y
A--ms-s
A--ms-f
Ncfs-n
Ncms-s
Ncmt
Ncfs-y
//nuisance, mischief; bother, trouble; difficulty//
//to bleach, whiten; peel, skin; shell; hull//
//to bleach, whiten; peel, skin; shell; hull//
//to bleach, whiten; peel, skin; shell; hull//
//to bleach, whiten; peel, skin; shell; hull//
//white//
//nuisance, mischief; bother, trouble; difficulty//
//white//
//white//
//white//
//whiteness//
//belote (card game) //
//belote (card game) //
//white, whiteness//
From Electronic Corpora to Online Dictionaries
89
Bilingual Bulgarian-Polish online dictionary
The Bulgarian-Polish online dictionary is being developed for experimental
purposes. A LDB provides the language material for the dictionary. For the
program realization of the web-based application the technologies Apache, MySQL,
PHP and JavaScript have been used; these are free technologies originally designed
for developing dynamic web pages with a lot of functionalities. The current version
of the Bulgaria-Polish online dictionary works optimally with Internet Explorer 6.0+
(Windows), and with Firefox 2.0.1+ (Windows, Linux). The website resolution is
1024/768 pixels.
The web-based application consists of two primary modules: an administrator
module and an end-user module.
The administrator module is intended for the person updating the dictionary,
and access to it is limited only to authorized users. The administrative module is
used to fill in the database and to offer user-friendly interface to the user who will
be responsible for the word management. This module recognizes two types of
users: (1) “super administrator”- who has all rights of adding, editing, deleting and
searching for words; adding, editing and deleting users and (2) “administrator”who has all rights except creating a new user and deleting an existing one.
The administrator module manages some main sections: a section for entering
a new word (see the example below), sections for searching for Bulgarian or Polish
words, a section where end-users report the missing words. The Help section serves
both the administrators and the end users.
Administrative panel - choosing the type of the word which will be added: a noun
The end-user module is the module, through which the end-user accesses the
information in the dictionary. The interface is bilingual, the user can choose the
input language (Bulgarian or Polish) and according to his/her choice, a virtual
Bulgarian or Polish keyboard is displayed. In this way the user can choose special
Bulgarian or Polish characters if they are not supported by his/her own keyboard.
There are three sections in this module: a section for translating a word, an
information section and a section for reporting a missing word. After making a
90
Ludmila Dimitrova
search for a word on the left site of the screen, a list of words is displayed starting
from the given entry. A click on any of these words in the list visualizes the
translation in the right frame. If we translate from Bulgarian to Polish, the whole
information saved in the LDB is displayed. When translating from Polish to
Bulgarian, only the Bulgarian headwords are visualized.
The program realizing the web-based application for representation of the
Bulgarian-Polish online dictionary allows the dictionary volume to be expanded by
adding new words, enriching the content of the dictionary entries from the LDB by
adding new examples for clarification of the meaning, etc.
Administrative panel –2nd step of adding the participle
Furthermore, the structures of the LDB and of the web-based application allow a
replacement of the Polish translations (texts) by texts in another language Lang.
Thus, the LDB and the web-based application can be useful for the development of
a new bilingual Bulgarian-Lang online dictionary.
6 Conclusion
In this paper I briefly presented the Bulgarian language resources which were
developed in the Mathematical Linguistics Department at the IMI-BAS in the
framework of some international projects.
Some possible directions for future work are: bringing the morphosyntactic
descriptions for verbal forms in line with the Bulgarian grammar and updatiung
Bulgarian MSDs and lexicon for MTE resources Version 4, extending bilingual
corpora, enriching bilingual LDBs with new entries and new languages, increasing
the number of headword classifiers, and increasing the speed of the search module
of the web-based application for representation of an online dictionary.
From Electronic Corpora to Online Dictionaries
91
Acknowledgement
I would like to thank all colleagues with whom I worked throughout the years for
the development of the Bulgarian multilingual resources: Lydia Sinapova and Kiril
Simov (Bulgarian Academy of Sciences, Sofia, Bulgaria), my colleagues from the
MTE and CONCEDE projects, V. Koseska-Toszewa (ISS-PAS), R. Garabík (ĽŠILSAS), R. Panova and R. Dutsova (my students from the MSc program Languages
and Multimedia Technologies of IMI-BAS – Veliko Tarnovo University).
References
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
Bulgarian Explanatory Dictionary. (1997). Л. Андрейчин и др. Български
тълковен речник. Четвърто издание. Допълнено и преработено от
Д. Попов. Издателство Наука и изкуство, София, 1997. (In Bulgarian).
Dimitrova, L., T. Erjavec, N. Ide, H. Kaalep, V. Petkevič, D. Tufiş. (1998).
Multext_East: Parallel and Comparable Corpora and Lexicons for Six
Central and Eastern European Languages. In Proceedings of the COLINGACL’98, pages 315-319, Montréal, Québec, Canada.
Dimitrova, L., V. Koseska–Toszewa. (2008). Some Problems in Multilingual
Digital Dictionaries. In: International Journal Études Cognitives. Vol. 8.
SOW, Warsaw. 2008, pages 237–254. ISSN 1641-9758.
Dimitrova, L., R. Panova, R. Dutsova. (2009). Lexical Database of the
Experimental Bulgarian-Polish online Dictionary. In: Metalanguage and
Encoding scheme Design for Digital Lexicography. Proceedings of the
MONDILEX Third Open Workshop, Bratislava, Slovak Republic, 15–16
April 2009, pages 36–47. ISBN 978-5-9900813-6-9.
Dimitrova, L., R. Pavlov, and K. Simov. (2002). The Bulgarian Dictionary in
Multilungual Data Bases. Cybernetics and Information Technologies. Vol. 2,
num. 2, pages 12–15.
Dimitrova, L., R. Pavlov, K. Simov, and L. Sinapova. (2005). Bulgarian
MTE Corpus – Structure and Content. Cybernetics and Information Technologies. Vol. 5, num. 1, pages 67–73.
Dimitrova, L., P. Rashkov. (2009). A New Version for Bulgarian
MULTEXT-East Morphosyntactic Specifications for Some Verbal Forms.
In: Organisation and Development of Digital Lexical Resources. Proceedings of the MONDILEX Second Open Workshop, Kiev, Ukraine, 2–4
February 2009, pages 30–37. ISBN 978-966-507-252-2.
Erjavec, E., R. Evans, N. Ide, A. Kilgarriff. (2000). The Concede model for
lexical databases. In Second International Conference on Language Resources and Evaluation, LREC’00, Athens, ELRA.
Ide Nancy. (1998). Corpus Encoding Standard: SGML guidelines for
encoding linguistic corpora. In First International Conference on Language
Resources and Evaluation, LREC’98, pages 463–470, Granada, ELRA.
http://www.cs.vassar.edu/CES/
92
[10]
[11]
[12]
Ludmila Dimitrova
Ide N. and J. Véronis. (1994). Multext (multilingual tools and corpora). In
Proceedings of the 15th CoLing, pages 90–96, Kyoto.
EAGLES 1996. Monachini, M. and Calzolari, N. Synopsis and Comparison
of Morphosyntactic Phenomena Encoded in Lexicons and Corpora: A
Common Proposal and Applications to European Languages. EAGLES
Report EAG-CLWG-MORPHSYN/R.
http://www.ilc.pi.cnr.it/EAGLES96/morphsyn/
MTE 2004. MULTEXT-East Morphosyntactic Specifications – version 3,
edition 10th May 2004.
Evaluating Grid Infrastructure
for Natural Language Processing
Radovan Garabík1, Jan Jona Javoršek2 , and Tomaž Erjavec2
1
L’. Štúr Institute of Linguistics, Slovak Academy of Sciences, Bratislava, Slovakia
[email protected]
2 Jožef Stefan Institute, Ljubljana, Slovenia
[email protected] [email protected]
Abstract. In this article we analyze common human language technology requirements and the possibility of implementing them using Grid infrastructure.
Different possibilities for the setup of an execution environment are treated and
the standard PKI based Grid security approach is explained, with an emphasis of
securing data access in a potentially untrustworthy environment. Two examples
of running unmodified NLP applications are presented.
1 Introduction
Increasing computing requirements for acquiring and processing large textual data-sets
and working with larger and larger corpora in Natural Language Processing (NLP) and
related disciplines together with ever increasing availability of computing resources
allow us to work on NLP algorithms and tasks that were impractical just a few years
ago. The core of the problem shifted from obtaining access to enough computation
power and from optimizing algorithms into developing efficient ways of allocating the
computing resources to various tasks and into finding efficient ways of dealing with
huge amounts of data. Since the most accessible computing environment moved from
large centralized supercomputers into the vast number of available servers connected
with the ubiquitous Internet, a new paradigm in computing emerged: massively parallelized algorithms running on widely distributed networks of interconnected computers.
One of the infrastructure approaches is the Grid network, which provides a complete environment for heavy-duty computational tasks, with working solutions for user
authentication, data storage, distributing load over the available resources, access control and the whole infrastructure for user management. First used mostly for computing
tasks in high energy physics, the Grid is nowadays used for tasks in several different
research areas. The use of Grid infrastructure for NLP has been previously discussed
with several proposals (cf. [19], [4], Neuroth et al. [14], [9], [10], [12], [11], [13],
reviewed in [6]).
In this article, we explore the idea of utilizing the Grid infrastructure for NLPrelated tasks. Not just the computational requirements, but also commonly useful features of shared data repositories and existing Grid security features are discussed. Since
the whole Grid environment runs exclusively on the GNU/Linux operating system
(although there are efforts underway to port the Grid software to other Unix-like platforms, such as the BSD family and Mac OS X), our point of view will deliberately be
94
Radovan Garabík, Jan Jona Javoršek, and Tomaž Erjavec
rather ‘GNU/Linux centric’, and the discussed software environment and tools will be
implicitly understood to be GNU/Linux specific (unless stated otherwise).
2 Legal issues
The actual deployment of Grid computing in the natural language processing area
(especially relevant for corpus linguistics) faces specific legal issues – the data being
processed are in majority of cases copyrighted, and the research institutions either have
very strict legal agreements governing the use of the data, or are operating entirely on
copyright law sections allowing scientific and research use of the data (fair use in the
U.S.A. jurisdiction, citation and educational use in many of the EU countries’ copyright
laws). The situation is somewhat similar to the problems the users of Grid computing in
health care systems – though in that case, metadata are the most sensitive and protected
part of the data-set, while in corpus linguistics the data (i.e. texts in the corpora) are
sensitive, but the metadata is usually freely accessible [17].
In any case, the research institution using the data for research most likely does
not have the right to distribute the data at all. If the contractual obligations prevent
the institution from physically copying the data beyond the premises of the institution,
it might be still advantageous to use the Grid infrastructure for computing clusters of
the institution itself, and use middleware functions to restrict data-replication to those
processing nodes and data storage elements physically located in the organization. This
way, the whole Grid can still be used for less sensitive tasks, or for post-processing
the results of operations on sensitive data (when the post-processing does not include
access to sensitive data), while at the same time the computing nodes will be available
as part of the whole Grid computing pool when they would be left idle otherwise.
While the actual uploading of the data to Grid-enabled storage is not to be considered a form of ‘distribution’ as long as no other person or organization is allowed to
get the data, it is nevertheless desirable to protect the data from casual snooping. For
one thing, an administrator of the Grid node where the data physically reside can get
access rather trivially; and while he or she is legally obliged not to misuse his access
(usually by rather strict agreements, in the case of European Grid infrastructure), a
measure of additional protection seems to be necessary – to avoid data leaking in case
the computer hosting the Grid node is compromised, unbeknown to the administrators.
We discuss security measures used in the Grid infrastructure in the following section.
3 Software environment in the Grid
The different implementations of Grid middleware all follow the same workflow: the
user has to provide a way to parallelize the computing task in a number of jobs that can
be run in parallel and encode the solution by providing a job specification, indicating the
data files required (using URIs), suitable computing environment (ABI, API, execution
environment), computing time and resources needed (wall-time, RAM, disk) and a way
to store the results.
Evaluating Grid Infrastructure for Natural Language Processing
95
This can be done manually, and a job is then submitted with a dedicated command,
or, alternatively, can be done with the help of a dedicated web application, usually
domain or experiment specific.
The system then takes care of selecting the appropriate Grid site and free worker
nodes, downloading the data files, making the pre-installed software (execution environment) available and starting the job script. A number of tools enable the user to
monitor the progression of the task, including its standard input and standard/error
output, working directory contents etc.
When the job finishes, it may upload the resulting data files to a Grid-enabled disk
storage, or, alternatively, the user can use command-line tools to download the data
from the working directory directly.3
To make the system work reliably and securely, a number of information, monitoring and accounting systems are part of the infrastructure. In addition, an advanced
security model is used to ensure resource protection and data integrity. As we have to
consider this system’s suitability for protection of copyrighted data-sets in linguistic
resources, a short overview is presented in the following subsection.
3.1 Security
Computing grids had to be very security-conscious from the very beginning, since the
very premise of a Grid network is, from the point of view of the site administrator, to
give external users access to the local computing infrastructure and, from the point of
view of Grid users, to entrust data and applications to untrusted, foreign sites.
Moreover, the basic requirement for a viable, scalable and sustainable security infrastructure in the context of large Grid networks has to be a robust solution with as few
single points of failure as possible to avoid failures of security services that could effect
negatively the availability of the whole infrastructure.4
Grid security has several components: (a) Authentication, a method of confirming
the identity of the user or organization behind an operation, is implemented on the
basis of the Public Key Infrastructure (PKI) and standard x509 digital certificates (with
a number of extensions to facilitate the use of PKI in the context of Grids). (b) Authorization is provided in the framework of virtual organizations (VOs), a mechanism
enabling Grid users all over the world to organize themselves according to research
topics and computing requirements, regardless of geographic constraints, and permitting sites to regulate the use of their resources according to user, discipline, software
requirements etc. (c) Monitoring and ticketing permits users and administrators to keep
track of infrastructure availability and to react to technical and security matters in a
timely fashion. (d) Accounting reports on the use of the infrastructure and enables the
community to regulate and enforce the use of the infrastructure.
Public Key Infrastructure. Public Key Infrastructure, first introduced to the general
public in the context of securing the web and enabling on-line shopping and banking,
has become the standard authentication model in many application domains. Defined
3
4
See [5] for architecture overview of the NorduGrid ARC middleware.
For an overview, see [8].
96
Radovan Garabík, Jan Jona Javoršek, and Tomaž Erjavec
by a number of Internet Drafts, RFCs and standards, PKI is a widely deployed and
evolving system.5
PKI is based on the property of asymmetric ciphers, where a different key is used for
encryption and decryption. This property allows the encryption key to be always kept
private and secret and the decryption key to be public, usually published with some
information about the owner of secret key in the form of a x509 digital certificate.
In PKI, such a digital certificate is used as the token of identification: it is issued
by a certification agency (CA) on the basis of an identification process (i.e. checking
legally acceptable personal ID documents in person). But the certificate is coupled with
a secret key that has been generated by the user requesting the certificate and is never
exposed to the CA. To issue a certificate, the CA now sets up information about the
entity (user, host or service) to be certified in accordance to the identification data
provided in a standard form called a Distinguished Name (DN, following a LDAPlike name scheme: CN = Joe User, OU = My Department, DO = Institute of
Dispersive Linguistics, DC = San Marino, and signs it with their own secret
key from the CA certificate.
This scheme ensures that nobody, not even the CA, can use the certificate (since
only the owner of the certificate possesses the secret key) and protects the information
in the certificate with the signature, produced with the CA’s own secret key.
To make the system work, CA certificates with public keys are published in a well
advertised manner (or shipped with software, such as. web browsers, Grid middleware
packages and GNU/Linux distributions). Recipient of a document or a connection that
uses a client certificate and is encrypted or signed with such a certificate can therefore
verify that the document or connection really was encrypted or signed by the said
certificate by decrypting it with the public key included in the certificate, and it can
verify the information in the certificate by checking the certificate with the CA public
key in the same manner.
A number of additional security measures are used in the Grid: CA secret keys are
kept in off-line systems or in dedicated certified hardware modules (hardware security
modules or HSM) while end-entity certificates are re-issued with new keys yearly or
kept in hardware security tokens. In addition, actual user certificates are never entrusted
to non-trusted entities: for almost all operations in the Grid, short-lived proxy certificates are used instead (described bellow).
Virtual organizations. While PKI provides authentication, a different system is needed
to provide authorization, i.e. to help decide if a given user, host or service is to be
allowed to carry out a specific task: use a specific resource or access specific data. In
the context of Grid computing infrastructure, this role is implemented in the framework
of virtual organizations (VOs).
A Virtual Organization serves two purposes: (a) As an organizational form, a VO
permits a number of researches from different organizations, usually geographically
dispersed, to collaborate and share tools, data and resources. (b) In the Grid security infrastructure, a VO provides means of regulating access to resources, i.e., a VO provides
authorization after authentication is provided by PKI.
5
See [2] for an extensive up-to-date overview of the relevant documents.
Evaluating Grid Infrastructure for Natural Language Processing
97
With this combination of roles, Virtual Organizations have proven themselves to be
most efficient in enabling a higher level of international collaboration and have permitted the European Grid network to foster new, faster development in many disciplines
by providing an unprecedented framework for international collaboration.
In practice, members of a research project or a discipline can set up a VO and
decide on its modes of operations and access to resources quite independently. They
have to decide what kind of tools the VO members will be using in the Grid, define the
data formats, prepare data repositories, develop execution environments with the tools
installed and set up a Virtual Organization Membership Service server (VOMS server)
to store authorization credentials.
Then some resources have to be made available to the community of VO members.
In practice, that means obtaining support of a number of Grid sites (organizations owning computing clusters partaking in the Grid) that have to configure their Grid middleware installations to include the new VOMS server in its authorization procedures and
to either install the execution environment (or, more realistically, environments) for the
VO or give access to some members of the VO so that they can perform the installation
and maintenance if the execution environment on the site themselves. Additionally, a
number of Grid storage elements (SE) has to be configured to allow the VO members
to access and store the data on their disk space.
Proxy certificates. With the VO and VO supporting Grid sites, a VO member can
submit Grid jobs and access VO-owned data using his certificate. This is implemented
in an indirect manner by means of Grid proxy certificates, as mentioned previously in
the discussion of PKI infrastructure.
Grid proxy certificates are (a) primarily used to permit a job to authenticate in the
name of the user spawning the job, without the requirement of direct user interactions
during the course of the job. This means that the proxy certificate must have the same
DN as the users’ certificate, but it has a different secret key which is not protected with
a pass-phrase that would require user interaction on the keyboard. Proxy certificates are
generated with a tool that uses the users’ certificate to sign the proxy (as if it were a
CA), thus confirming that the proxy was indeed generated by the user. In addition, gird
proxy certificates are protected with file permissions and are always short-lived (from
several hours to a few weeks) to mitigate the risk of the unprotected secret key.
To interact with the VO authorization system, the user generates a VOMS Grid
proxy certificate that obtains special certificate extensions from the VOMS server and
incorporate them in the proxy certificate. These extensions encode VO group and role
attributes of the user and are themselves signed by the VOMS server with its service
certificate, using the PKI infrastructure’s authentication facilities to implement an authorization layer.
In this manner, a job can obtain authorization to use computing resources and data
simply by providing a suitable VOMS proxy certificate. Its attributes are recognized by
the Grid manager servers that provide it with to data storage (storage resource managers,
SRM) and other resources.
As and additional level of security, Grid managers assign each job a temporarily
unique user ID in the underlying operating system mapped from its active VO role in
98
Radovan Garabík, Jan Jona Javoršek, and Tomaž Erjavec
such a way that no jobs with different roles (and therefore potentially different access
permissions) can share access on the underlying implementation.
In this way the system implements fine-grained control over the use of Grid resources and data without any reliance on the availability of authentication and authorization servers, thus avoiding a single point of failure that would have a significant
impact on the scalability of the system.
Data Protection. Using these security components, additional measures of data protection can be implemented when necessary. In the context of NLP, such a measure is of
critical importance, since most of the data-sets in corpus linguistics contain copyrighted
texts that need to be protected.
To solve this problem, the corpus data has to be suitably protected where it is
permanently stored. Therefore we propose to store the corpus data in encrypted form
in a dedicated storage element and set up the access authorization in such a way that
access is restricted to VO users who belong in a VO group of users who signed the
necessary legal agreements to access the data. Furthermore, we propose that the data
is transferred to the untrusted environment of Grid worker nodes, where jobs perform
their computations, in the encrypted form and that the decryption keys are issued to the
jobs protected with asymmetric encryption decryptable only by the job’s Grid proxy
keys so that only the jobs can access the keys and decrypt the data.
In this manner, access and decryption is regulated with the authorization of embedded VOMS attributes in the proxy certificate without any additional authorization steps,
while the data is never shipped or stored in unencrypted form.
If the tools used by the job have to store temporary files on disk, these are protected
from other processes (with the exception of system administrators, who are already
bound by strong agreements pertaining to data security on the Grid) and are in addition
of short-lived nature.
There exist different implementations of the system described. The simplest form
involves the use of a decryption filter in the job script and is rather simple to deploy. A
more flexible solution, based on CryptoSRM (cryptographic storage resource manager)
and Hydra Key Storage (a distributed fragmented encryption key storage system) is
described in [17].
4 General requirements of NLP related tasks
Contemporary NLP tasks are rather varied; some of them require a lot of “pure” computing power, but many tasks, especially in the area of corpus linguistics, merely process large data files. From the software point of view, the tools used cannot be more
diverse – they are often programmed in typical computer languages, like C or C++,
but a lot of data processing is done in scripting languages, such as Perl or Python, and
Java is increasingly popular, and more often than not, one specific task uses several
different tools bound by short programs written in a shell script. The use of (high level)
scripting languages even for the computing intensive tasks means that the analysis is
less effective than it could be, but the ease of creating and maintaining the tools more
than outweights this particular disadvantage. From this follows than the tools are often
Evaluating Grid Infrastructure for Natural Language Processing
99
fragile and require a specific environment, which sometimes means that even using a
different GNU/Linux distribution that the one the software has been developed on can
be a major problem.
The Grid environment, due to its initial connection with the use in High Energy
Physics, predominately uses Scientific Linux CERN distribution (SLC) version 4 for
the job computing environment (with a changeover to version 5 currently in progress).
The ideal solution would be of course to put all the necessary NLP software into the
execution environment (which is available at each of the computing nodes) and use
the standard distribution. It is, however, sometimes much more convenient to use an
operating system environment more suitable for the users and their tools. There are
two possible solutions: to run under a chroot environment or to use virtualization. Both
options are discussed below.
4.1 Userspace and full virtualization
Chroot is a UNIX system that changes the effective root of the filesystem for the process
and its children. The basic usage for chroot is twofold: it can be used to restrict untrusted
(or potentially dangerous) processes from accessing the rest of the filesystem, or it
can be used to run processes in a different filesystem environment (different filesystem
layout with different system executables and dynamic libraries). It should be noted that
chroot does not offer true virtualization since isolation from the host system is not complete – in particular, system kernel, networking subsystem and process management are
shared with the host system, so that the processes in the chroot environment cannot bind
to sockets that are used on the host system (and vice versa), and if process management
is to be possible in a chroot environment, the proc filesystem has to be mounted inside
chroot environment, enabling the guest to access the information about host processes6 .
On the other side of the spectrum, there are complete virtualization solutions, emulating the guest system. These can emulate the CPU completely in software (approach
commonly used in emulating vintage computers on modern operating systems, or when
a computer platform switches the architecture), or run the guest machine natively, trapping and emulating only privileged or unimplemented instructions. Modern computer
architectures usually offer dedicated hardware features to facilitate the implementation
of virtual machines7 , some mainframe architectures even offer complete, seamless virtualization in hardware.
Then there are several different approaches that lie somewhere in between those
two extremes, ranging from (a) paravirtualization, which requires cooperation from the
guest operating system kernel (in order to achieve negligible performance loss due to
the virtualization), used e.g. by the XEN virtualization solution; to (b) compartmentalization (i.e Linux virtual servers and OpenVZ), which divides the host operating system
into different compartments with completely separated processes, network access and
6
7
This does not matter as much as it seems as long as the chrooted processed run under different
PID from the host ones, because a non-root user cannot affect other processes, and a chrooted
superuser can break out of the chroot anyway.
Until rather recently (before the introduction of VT-x and AMD-V), such features were not
available in common Intel-compatible off-the-shelf computers.
100
Radovan Garabík, Jan Jona Javoršek, and Tomaž Erjavec
filesystems but sharing the same kernel; to vanilla kernel namespace support, which
only separates user and process management (slightly extending chroot separation).
The virtualization techniques mentioned differ on performance impact [15] – ranging from none at all in case of a simple chroot or chroot with namespaces, over very little
for OpenVZ-like compartmentalization to a more significant one for full virtualization.
The specific areas of impact vary, too – while the raw CPU performance rarely decreases
by more than a few percent (with the exception of complete software emulation of the
guest architecture), I/O penalties are sometimes severe.
From our point of view, the best way to use the specific software is to install it inside
a runtime environment which is made available to the jobs when submitted to the Grid.
This is directly supported by the Grid infrastructure and requires no additional steps or
privileges. However, at this time this requires a significant effort, since all the tools and
their dependencies have to be compiled (or installed in a non-standard location inside
the runtime environemnt) on the standard SLC distribution, which can be a problematic
if the software has many external dependencies.
Installing a chroot environment, on the other hand, enables us to avoid porting the
software to the SLC distribution – inside the chroot, we can install any reasonably
standard GNU/Linux distribution and any necessary software packages. In addition,
many of the commonly used distributions already have support for (at least partial)
installation inside a chroot environment built in. But in the context of Grid infrastructure this solution has a significant disadvantage, since it requires support from the
cluster administrator since chroot environments are not a standard feature of the Grid
environment.
Using a complete virtual machine allows us to run a complete GNU/Linux distribution, with completely separate networking support and user management, including
the ability to run processes with superuser privileges, and the ability to use filesystems
otherwise not supported by the host system8 . But the main advantage is the possibility
to run completely different operating system9 . However, installing and using virtual
machines requires not just administrator cooperation, but often also nonstandard host
operating system extensions (such as special kernel modules). One of the more interesting virtualization systems in this context is User Mode Linux, which does not require
any special host support, runs as an ordinary user process and provides a complete guest
Linux kernel environment. Unfortunately, guest environment in this case suffers from
a big I/O performance degradation, which can be a noticeable problem when dealing
with very large corpus data.
While there is significant research in the use of different kinds of virtualization in
the context of Grid technologies, this is not a wide spread feature at this time. We have
been able to use clusters with full support for chroot environemnts, but we realize that
for quick adoption and widespread use of Grid computing in NLP, porting of tools to
the most often supported environment, i.e. SLC, will be necessary.
8
9
Such as encrypted filesystems.
R
R
family of
Windows
Therefore we can use e.g. the tools available only for Microsoft
operating system, if we can get around their mostly point-and-click nature and run them
noninteractively.
Evaluating Grid Infrastructure for Natural Language Processing
101
5 Proof of concept
This section presents and experiment in using the Grid to execute two NLP tasks for
the Slovak language. The first subsection introduces TectoMT, a machine translation
system, the second morphosyntactic tagging with morče, and the third gives the experimental usage of the two systems on the Grid.
5.1 TectoMT
TectoMT [20] is a software framework aimed at machine translation at the tectogrammatical level of analysis. The system is modular – the framework itself consists of
many independent modules (blocks in TectoMT terminology), each implementing one
specific, independent NLP-related task. Each of the blocks is a Perl module that interacts with the system using a single, uniform interface. However, sometimes the module
serves only as a wrapper for the underlying implementation in another programming
language. The tectogrammatical annotation and consequently the TectoMT framework
primarily stores linguistic data in its own format, called TMT. TMT is an XML-based
format, designed as a schema of the Prague Markup Language (PML)10 [16]. Nevertheless, its blocks are by no means obliged to use this format.
TectoMT has been developed with modern Linux systems in mind, and as such
its installation requirements are easily met by any contemporary Linux distribution.
It should be noted that TectoMT, being written mostly in Perl, depends on many external Perl modules and its installation scripts are intelligent enough to automatically
download and install any missing dependencies; this, however, circumvents standard
distribution packaging systems, therefore it is better to install all the necessary packages
with the packaging system tools before attempting to install TectoMT. There are also
some C language modules that are not compiled by default, but have to be compiled
separately inside the TectoMT installation source tree.
TectoMT also has some built-in capabilities for parallelization of its tasks, using the
Sun Grid Engine – it is possible to adapt the Sun Grid Engine batch software to various
Grid middlewares [3], but TectoMT can be run on the Grid system directly without
relying on its internal parallelization possibilities, if the user takes care of splitting the
input data into appropriate chunks for parallel processing.
5.2 Tagging a corpus
Morphosyntactic tagging of the Slovak National Corpus consists of two steps. The first
performs morphosyntactic analysis, where each word in the input texts is assigned a
set of possible morphosyntactic tags. This step essentially consists of looking up the
possibilities of lemma/tag combinations in a constant database table using the wordform
as a key, with an additional step for unknown words, where the list of possible tags is
derived from the similarities of word endings to the ones present in the database tables.
The software is implemented in the Python programming language and is actually quite
fast, since the core of the task consists simply of a look-up in the possibilities in the
10
Not to be confused with the Physical Markup Language.
102
Radovan Garabík, Jan Jona Javoršek, and Tomaž Erjavec
tables, and most of the CPU work is spent on I/O operations, parsing the input file and
assembling the output. On a reasonably recent hardware (Intel Xeon 2.33 GHz CPU) it
is able to process over 10 000 words per second. It can also parallelize easily, since the
words can be analyzed independently of each other.
The second step is disambiguation, where each word is assigned a unique lemma
and a morphosyntactic tag out of the possibilities assigned in the first step. For disambiguation, we use morče, an averaged perceptron model originally used for the
Czech language tagging [18], re-trained on the Slovak manually annotated corpus.
Disambiguation is much slower that the morphology analysis, its average speed reaches
only about 300 words per second. Parallelization at the application level is also not
possible without some redesign of the morče itself, but the nature of tagging makes it
easy to split the input data into as many chunks as we want and run morče instantiations
in parallel.
Given the speed differences between morphology analysis and disambiguation, we
can safely consider the morphology analysis execution time negligible and design the
whole tagging to be done in one step, without the need to parallelize the morphology
analysis process while the disambiguation is to be run in parallel.
5.3 Installation and usage
As our GNU/Linux distribution of choice is Debian11, we did all the testing on the
Squeeze (testing) Debian distribution, which is a “moving target” distribution, meant
for users that want newer version of the distribution and included packages, but do not
want to deal with (potentially) broken bleeding edge packages from the unstable Debian
repositories. To summarize, a package will get into testing if it has no release-critical
bugs, has spent several days in the unstable repository and its inclusion in testing will
not break other packages. We used testing deliberately, because it is advantageous to use
new versions of the required packages which will not become obsolete in near future,
even if the packages in testing repositories will be rather quickly replaced by still newer
versions.
Debian has a standard method for installing the base system into a chroot environment, implemented by a tool, called debootstrap[1].
Once the chroot environment was created, we proceeded and installed TectoMT Perl
dependencies into the chroot system. Finally, we ran the TectoMT installer. Similarly
we installed the Slovak version of the morče software from within the chroot. The
installation process was unremarkable comparing to a regular installation.
After the chroot environment had been prepared and installed on the Grid servers,
we were able to submit batch jobs with either TectoMT or the disambiguation tasks. We
used the Grid infrastructure to morphosyntactically tag and analyse the spoken Slovak
language corpus [7]. We divided the corpus (434 676 words) into 11 approximately
equally sized chunks and submitted them to be tagged in 11 independent batch jobs,
then joined the results. The run times ranged from 103 to 236 seconds, with the average
time around 141 seconds per job12 . Even including the overhead spent on submitting
11
12
http://www.debian.org
Times mentioned are wall times, i.e. total time elapsed from the beginning to the end of the
task, not including the time spent waiting in the job queue or downloading the data files.
Evaluating Grid Infrastructure for Natural Language Processing
103
and waiting for the job to start, this is a significant reduction of the total time needed to
tag the corpus. In fact, as far as the typical Grid tasks are concerned, our jobs were very
short, and we could achieve less overhead by running longer jobs (a typical Grid jobs
takes from 2 hours to more than a day). Such a s setup will actually be typical when
tagging a bigger, representative text corpus.
6 Conclusion
We demonstrated the possibility to use the Grid infrastructure for NLP related task.
From user point of view, a Grid computer behaves like any ordinary workstation running
Scientific Linux CERN distribution; however, in this article we discussed different
methods of using a custom GNU/Linux environment (or even another operating system) to better support the tools needed by the user. We set up and tested a chroot
environment running Debian GNU/Linux Squeeze distribution. Installing and running
TectoMT framework inside the chrooted environment was straightforward; similarly we
experienced no obstacles in installing and using the Slovak language morphology tagger
– we therefore do not expect any problems in deploying all kinds of “well behaved”
Unix (Linux) based software.
However, in order to truly exploit the Grid potential, we envisage a scheme where
the linguistic data (especially text corpora) are stored on the Grid infrastructure as well,
and the existing Grid access control infrastructure is extended in order to be provide
secure access to the data to third parties interested in accessing the data in such a way
that all the limitations and conditions arising from the copyright law and other binding
agreements are met.
In this way, we hope that the Grid infrastructure will soon become available to
researches in computational linguistics and, by multiplying the computing resources
available, will speed up linguistic research tasks and enable us to develop new algorithms, research methods and tools.
References
[1] Installing new Debian systems with debootstrap.
http://www.debian-administration.org/articles/426.
Retrieved 2009-10-05.
[2] Public-Key Infrastructure (X.509) (pkix).
http://www.ietf.org/dyn/wg/charter/pkix-charter.html.
Retrieved 2009-10-10.
[3] Borges, G., David, M., Gomes, J., Fernandez, C., Lopez Cacheiro, J., Rey Mayo,
P., Simon Garcia, A., Kant, D., & Sephton, K. (2007). Sun Grid Engine, a new
scheduler for EGEE middleware. In IBERGRID – Iberian Grid Infrastructure
Conference.
[4] Carroll, J., Evans, R., & Klein, E. (2005). Supporting Text Mining for e-Science:
the challenges for Grid-enabled Natural Language Processing. In Proceedings of
the UK e-Science All Hands Meeting.
104
Radovan Garabík, Jan Jona Javoršek, and Tomaž Erjavec
[5] Ellert, M., Gronager, M., Konstantinov, A., Konya, B., Lindemann, J., Livenson,
I., Nielsen, J., Niinimaki, M., Smirnova, O., & Waananen, A. (2007). Advanced
Resource Connector middleware for lightweight computational Grids. Future
Generation Computer Systems, 23(2), 219–240.
[6] Erjavec, T. & Javoršek, J. J. (2008). Grid Infrastructure Requirements for Supporting Research Activities in Digital Lexicography. In Mondilex: Lexicographic
Tools and Techniques, (pp. 5–13). IITP RAS.
[7] Garabík, R. & Rusko, M. (2007). Corpus of Spoken Slovak Language. In Levická, J. & Garabík, R. (Eds.), Computer Treatment of Slavic and East European
Languages, Brno. Tribun.
[8] Laccetti, G. & Schmid, G. (2007). A framework model for grid security. Future
Generation Computer Systems, 23(5), 702–713.
[9] Luís, T., Martins de Matos, D., Paulo, S., & Ribeiro, R. D. (2008). Natural
Language Engineering on a Computational Grid (NLE-GRID) T5 – Performance
Experiments. Technical Report 35 / 2008, INESC-ID, Lisboa.
[10] Martins de Matos, D., Luís, T., & Ribeiro, R. D. (2008). Natural Language
Engineering on a Computational Grid (NLE-GRID) T1 – Architectural Model.
Technical Report 38 / 2008, INESC-ID, Lisboa.
[11] Martins de Matos, D. & Ribeiro, R. D. (2008). Natural Language Engineering on a
Computational Grid (NLE-GRID) T2h – Encapsulation of Reusable Components:
Lexicon Repository and Server, January 2008. Technical Report 32 / 2008,
INESC-ID, Lisboa.
[12] Martins de Matos, D., Ribeiro, R. D., Paulo, S., Batista, F., Coheur, L., & Pardal,
J. P. (2008). Natural Language Engineering on a Computational Grid (NLEGRID) T2 – Encapsulation of Reusable Components. Technical Report 31 / 2008,
INESC-ID, Lisboa.
[13] Marujo, L., Lin, W., & Martins de Matos, D. (2008). Natural Language Engineering on a Computational Grid (NLE-GRID) T3 – Multi-Component Application
Builder. Technical Report 33 / 2008, INESC-ID, Lisboa.
[14] Neuroth, H., Kerzel, M., & Gentzsch, W. (Eds.). (2007). German Grid Initiative
D-Grid. Universitätsverlag Göttingen.
[15] Padala, P., Zhu, X., Wang, Z., Singhal, S., & Shin, K. G. (2007). Performance
evaluation of virtualization technologies for server consolidation. Technical
report, HP Laboratories.
[16] Pajas, P. & Štěpánek, J. (2006). XML-based representation of multi-layered
annotation in the PDT 2.0. In Hinrichs, R. E., Ide, N., Palmer, M., & Pustejovsky,
J. (Eds.), Proceedings of the LREC Workshop on Merging and Layering Linguistic
Information (LREC 2006), (pp. 40–47)., Genova, Italy.
[17] Santos, N. & Koblitz, B. (2008). Security in distributed metadata catalogues.
Concurrency and Computation: Practice and Experience, 20(17), 1995–2007.
[18] Spoustová, D., Hajič, J., Raab, J., & Spousta, M. (2009). Semi-supervised training
for the averaged perceptron POS tagger. In EACL ’09: Proceedings of the
12th Conference of the European Chapter of the Association for Computational
Linguistics, (pp. 763–771)., Morristown, NJ, USA. Association for Computational
Linguistics.
Evaluating Grid Infrastructure for Natural Language Processing
105
[19] Tamburini, F. (2004). Building distributed language resources by grid computing.
In Proc. of the 4th International Language Resources and Evaluation Conference,
(pp. 1217–1220).
[20] Žabokrtský, Z., Ptáček, J., & Pajas, P. (2008). TectoMT: Highly modular MT
system with tectogrammatics used as transfer layer. In ACL 2008 WMT: Proceedings of the Third Workshop on Statistical Machine Translation, (pp. 167–170).,
Columbus, OH, USA. Association for Computational Linguistics.
Synset Building Based on Online Resources
Ján Genči
Department of Computers and Informatics
Technical University of Košice, Slovakia
Abstract. A WordNet and EuroWordNet are two well-known projects the
aim of which is to build so-called synsets (set of synonyms) in order to
document meanings of the words. The purpose of the WordNet project was to
build English database. The purpose of the EuroWordNet project was to
develop “national” versions of WordNet databases including so-called
“interlingua index” for seven European languages. Since Slovak language
was not included in the original EuroWordNet project, there is an initiative to
find the way how to build relevant Slovak synsets and connect them to the
Interlingua Index. The paper describes several approaches which can be used
to build Slovak synset equivalents using (mainly) on-line resources
(resources available on the Internet).
1 Introduction
A WordNet [1] and EuroWordNet [2] are two well-known projects the aim of
which is to build so-called synsets (set of synonyms) in order to document meaning
of the words. WordNet specifies semantic network of English words. Project ontology for particular sensets of words defines (beside others):
•
More specific senses (hyponyms)
•
More general senses (heperonyms)
•
Senses of which semantic entity consists (meronyms)
•
Explanation of sense by:
• set of synonyms or simple word
• description
• example
Presented approach means standard way of sense specification in mono- and
bilingual dictionaries. The purpose of the WordNet project was to build English
database. The EuroWordNet [2] project developed “national” versions of WordNet
databases including so-called “interlingua index” for seven European languages
(Dutch, Italian, Spanish, German, French, Czech and Estonian). The Interlingua
Index links the EuroWordNet synsets to the original WordNet synsets.
WordNet and EuroWordNet proved themselves as valuable source of information not only for linguists but also for computer scientists working in the area of
text processing. Computer linguistics can use both networks of senses and senses
itself (specified as set of synonyms) in the process of solving different tasks. That is
a reason why scientists from many other countries are interested in having their
“national” version of WordNet. As an example we can state BalkaNet project,
purpose of which was to “enrich” EuroWordNet database by some other languages
(Greek, Turkish, Romanian, Bulgarian and Serbian) [3].
Synset Building Based on Online Resources
107
Since Slovak language was not included in the original EuroWordNet project
and there were no (mainly financial) sources to build Slovak compilation of
WordNet, there is question if it is possible to automate process of finding relevant
Slovak synsets and connect them to the synsets in other languages and moreover
link then to Interlingua Index.
2 Sources of synonyms (synsets)
The sets of synonyms for particular language can be acquired by several ways. The
first, “direct” approach is to use sources which specify synonyms explicitly. Other
ways, “indirect”, are based on algorithms for synset extraction from available
sources. They can use either explicit synset specification or combination of several
sources.
2.1 “Direct” sources of synsets
As explicit sources of synsets can be regarded:
•
WordNet [1] – English synsets
•
EuroWoirdNet [2] – Dutch, Italian, Spanish, German, French, Czech and
Estonian
•
BalkaNet [3] – Greek, Turkish, Romanian, Bulgarian and Serbian
•
Thesauri for particular languages
All resources stated above contain synset specified in the direct way. As another,
quite straightforward source of synsets we can regard monolingual or translation
dictionaries (bilingual or multilingual). Monolingual dictionaries specify senses of
words in a specific way usually as list of senses. Translation dictionaries use
notation very similar to notation used in the WordNet (or vice versa) – list of
synonyms, description of sense and examples. Public availability of presented
sources varied. While WordNet and on-line translation dictionaries are publicly
available, EuroWordNet and BalkaNet are closed project and they are available by
special permission only.
2.2 “Indirect” sources of synsets
Series of synonyms can be found by many different ways using various sources.
While direct sources of synsets can be used as source of synset for particular
language at the same time they can serve as an indirect source for synsets for other
languages. That was a case of the WordNet, which was used as source of synsets in
the EuroWordnet project.
Translation dictionaries we regard as direct source of synsets. However, on-line
translation dictionaries represent special case. They usually do not specify
translation of particular senses of the word but state the list of all translations only.
That is a reason, why we regard them as indirect sources.
108
Ján Genči
The paper presents several ways of producing Slovak synsets linked to synsets in
other languages. We present approaches based on the WordNet, and on-line
translation dictionaries. More complex approach is also mentioned.
3 Formal model of synset translation
It is clear that translation of the words between languages have to be based on the
senses of the word. We try to formalize this process for better understanding.
Def. Let Wi is a word, expressing sense V. Synset S (set of synonyms) is a set of
words, expressing sense V. Formally:
S={Wi | 0<i<n},
where n is a number of synset members.
Def. Let Vi is a sense, expressed by a word W. Homset H (set of homonyms) is a set
of all senses expressed by word W. Formally:
H={Vi | 0<i<n},
where n is a number of senses.
Let us take word W. Generally, word W can have several senses, what means it
is a member of homset H, with r members. Every sense of homeset H, can be
expressed by synset Si (1≤i≤r) (Fig. 1).
S1 = {w11, w12, …,W, …,w1s }
S2 = {w21, w22, …,W, …,w2s }
…
Sr = {wr1, wr2, …,W, …,wrs }
1
2
r
Fig. 1. Senses of word W specified by corresponding synsets
Example
To illustrate presented approach we use on-line version of WordNet [1]. Let us
take word accident. Result provided by the WordNet is presented on the Fig. 2.
We can see that homset (word) accident has two members (i.e. has two senses):
1. an unfortunate mishap; especially one causing damage or injury
2. anything that happens suddenly or by chance without an apparent cause
The first one is a synset consisting of a single word, second one is consisting of
several words (accident, stroke, fortuity, chance event).
Def. Let TLL (W ) is an operation of translation of word W from language L1 to
language L2.
2
1
L
Translation TL (W ) can be semantically expressed according Fig. 3.
2
1
Synset Building Based on Online Resources
109
Fig. 2. WordNet result for word accident
L1
L1
L1
L1
L1
L2
L2
L2
L1
L2
L2
2
L2
21
L2
22
L1
2i
L2
2t2
2
2
2
1
2
S 1 ={w 11 , w12 ,... , W ,... , w 1s }V 1  S 1 ={w11 ,w 12 ,... , w 1i ,... , w 1t }
1
L1
2
L1
21
L1
22
L1
1
1
1
L1
2s2
S ={w , w , ... ,W , ... , w }V 2  S ={w , w , ... , w , ... , w }
...
L
L
L
L
L
L
L
L
L
L
S r ={w r1 , wr2 , ... ,W , ... , w rs }V r  S r ={w r1 , w r2 ,... ,w ri , ... , wrt }
1
1
1
r
r
Fig. 3. Semantics of translation of word W
The result of the operation TLL (W ) we can express as
2
1
L2
L2
L2
{S 1 , S 2 ,... , S r
It is clear that word W is member of each synset SiL . Every sense represented by
synset in the source language is translated to destination language as a set of words
– synset which represents translated sense in the destination language.
1
Using two-way translation (forward and back) for given word W we can try to
„reconstruct“ particular synsets using translation dictionary according to following
algorithm:
L2
L2
L2
L2
1. get T L W ={S 1 , S 2 ,... , S r }
1
L2
2. for each w i ∈S i
L
L
L
L
a R i =T L w i ={S 1 , S 2 ,... , S q } // get back translation
1
L1
b S i =
ti
1
1
1
2
∩R
k =1
k
Presented algorithm gives us possibility to build relevant (linked) synsets in both
languages.
110
Ján Genči
4 Synset building – theoretical view
On the base of presented model we have implemented or plan to implement several
approaches to build synsets for Slovak language:
1. using synsets from the WordNet database
2. using bilingual on-line translation dictionaries
3. using (publicly) available thesauri
4.1 Using WordNet database
As stated above the WordNet specifies senses of words by synsets. Moreover, it
specifies relationships between word senses – more general and more specific word
senses. To determine synsets of destination language (Slovak language in our case)
we are using two approaches:
1. intersection of translation of particular WordNet synsets words
2. translation of related words in the WordNet network (hyponyms and
hyperonyms)
The first approach we use in the case if particular synset is defined by multiword
synset. The second one is used in the case of single word synset.
We try to illustrate our approach on the case of word computer. According to
WordNet, this word has two noun senses:
1. computer, computing machine, computing device, data processor, electronic
computer, information processing system (a machine for performing
calculations automatically);
2. calculator, reckoner, figurer, estimator, computer (an expert at calculation
(or at operating calculating machines));
Result of determination of Slovak equivalent by intersection of translation of
English synset members is presented in the Fig. 4.
Fig. 4. Building Slovak synset based on WordNet data
Using specified approach for several languages it is possible to build linked
multilanguage synsets.
Synset Building Based on Online Resources
111
4.2 Using on-line translation dictionaries
Classical printed dictionaries specify senses of word as lists of words in the
destination language. For many reasons this lists are not “full” synsets. However,
even not complete synsets can be used to determine the translated sense of word
using similar approach we used in the WordNet case. The only problem is on-line
availability of such source.
In the contrary to printed version of dictionaries, corresponding on-line
dictionaries usually lack information regarding sense specification. For example,
from the first ten results of Google search condition “slovnik site:sk”, the only one
dictionary [4] provides results in the structured form. Other ones provide only
unstructured translation of given word. However, even such approach gives us
opportunity, using forward and back translation, to determine some synsets in both
languages.
Let us take Slovak word jednotka. Translating this word from Slovak to English
using dictionary FEEHA [5] we got translation specified by tab.1. (results for
multiword senses are eliminated). Intersecting backward translations (with simultaneous elimination of original word jednotka) we got result according Tab. 2.
Adding word jednotka to computed synsets we get results according Tab 3.
Translation to
English
Backward translation to Slovak
drive
jazda, pohon, nápor, vychádzka, cesta, vozovka, budenie, jednotka,
mechanika, zariadenie, podnikavosť, motivácia, prevod, nábor,
riadenie, výjazd
element
člen, častica, živel, súčiastka, časť, článok, prvok, element,
jednotka, zložka
entity
podstata, jednotka, celok, bytosť, bytie, teleso, objekt, entita,
existencia, predmet
force
sila, úsilie, vplyv, moc, vodopád, tlak, tiaž, zbor, oddiel, jednotka,
platnosť, účinnosť, pôsobnosť, násilie, donútenie, zmysel, tvárnik,
zasahovanie, zásah, útvar
handler
zariadenie, jednotka, manažér
item
poznámka, prvok, bod, vec, jednotka, článok, odstavec, číslo,
položka, člen, slovo, rovnako, detail, zápis, odsek
scaler
reduktor, čítač
timer
časovač, rozdeľovač, krokovač, hodiny, jednotka, čas
unit
jednotka, celok, oddiel, útvar, jednička, jedna, zväz, diel, agregát,
stroj, prístroj, zariadenie, lístok, ročník, kurz, terminál, kus,
skupina, zložka, blok, člen, spojenie
unity
zhoda, súdržnosť, zomknutosť, jednotka, jednota
Table 1. Forward and backward translation of word jednotka
112
Ján Genči
#
Intersection based on
Slovak word
English word whose translation contains
specified Slovak word
1
celok
entity, unit
2
zariadenie
handler, drive, unit
3
člen
element, item, unit
4
útvar, oddiel
force, unit
5
prvok, článok
element, item
Table 2. Result of backward Slovak translation
#
Slovak synset
English synset
1
jednotka, celok
entity, unit
2
jednotka, zariadenie
handler, drive, unit
3
jednotka, člen
element, item, unit
4
jednotka, útvar, oddiel
force, unit
5
jednotka, prvok, článok
element, item
Table 3. Linking Slovak and English synsets
4.3 Using thesauri
Thesauri, as mentioned above, are natural source of synsets. We can use them
similarly as we used WordNet database – taking the set of synonyms, translating
each word and intersect particular translations. However, different approach can be
used if we have thesauri for both languages. In this case we can try to find
corresponding synsets and link them together.
4 Conclusion
At the moment of manuscript preparation we have implemented
two approaches:
1. building synsets based on WordNet using translation dictionaries [6];
2. building synsets using on-line translation dictionaries only [7].
Experiments performed by implemented approaches showed that none of them
is perfect. Each of them produced very good result for some words but at the same
time very poor results for other words. Results can be affected by several factors:
•
word used to specify synsets are usually not true synonyms; they often
represent close senses only;
•
translation dictionaries do not contain all synsets;
Synset Building Based on Online Resources
•
•
•
113
different authors (editors) presents different subsets of word constituting
synsets in their dictionaries;
no dictionary specifies all senses of a word (for most of the words);
two-way dictionaries (especially printed versions) are not symmetrical (we
can not be sure that for translation wL₁ → wL₂ we will find translation
wL₂ → wL₁
In the future we plan to integrate presented approaches in the common single
approach which will be used for building relevant linked synsets.
References
[1] WordNet. A lexical database for the English language.
http://wordnet.princeton.edu/
[2] EuroWordNet. A multilingual database with wordnets for several European
languages. http://www.illc.uva.nl/EuroWordNet/
[3] Balkan WordNet (BALKANET).
http://www.ist-world.org/ProjectDetails.aspx?
ProjectId=a137b147dead4b75b11d4d8da46e7767
[4] Online Dictionary. Slovensko-maďarský slovník
http://slovnik.agx.sk/dictionary_main_sk_hu.php
[5] Translation dictionary FEEHA.
http://www.feeha.sk/translator.php
[6] Lapoš P.: Overenie budovania EuroWordNet synsetov na báze on-line slovníkov. Diploma thesis. KPI FEI TU Košice. 2005. In Slovak.
[7] Sudynová M.: Nástroj na generovanie slovníkových záznamov. Diploma
thesis. KPI FEI TU v Košiciach. 2006. In Slovak.
Shallow Ontology Based on VerbaLex
Marek Grác
NLP Centre, Faculty of Informatics, Masaryk University
Botanická 68a, 602 00 Brno, Czech Republic
Abstract. Ontologies have proven to be a useful resource in natural language
processing. In this paper, we introduce basic ideas of a shallow ontology named
Sholva. This ontology is based on VerbaLex, a database of verb valencies, where
each valency pointer also contains a pointer into EuroWordnet. We focused our
effort on building ontology which would help us in solving real problems in
syntactic analysis, word sense disambiguation and machine translation.
1
Introduction
The information around us is increasing with every second. One is no longer able to read
and process it in a lifetime. Processing of natural language on computers can help us in
finding the most relevant pieces. Various resources are used for this purpose. Semantic
networks are among the most popular formalisms for knowledge representation. Like
other networks, they consists of nodes and links. For English we have a number of possibilities from domain oriented to general ones like Princeton WordNet [8], for Czech one
big general semantic network EuroWordNet [12]. It is very rich and complex network
but unfortunately only few applications use its potential. Creation of similar resources
for other close languages such as Slovak is very difficult and also time-consuming.
Our goal is to create a simpler ontology which will be easy to create and use primarily in our existing applications. This application-driven approach should help us to
avoid creating a perfect complex ontology by providing us with a simple one instead,
which can be used in various projects right now. Simpler ontology should also help
us to create similar resources for other languages and take advantage of it in machine
translation. Such project can reuse many existing components that were created for
different purposes and projects.
2
WordNet ontologies
A lexicon with information about how words are used and what they mean is a necessary component for any application working with natural language. Ontologies are one
of the resources that can provide enough information for those. Ontology is a formal
representation of a set of concepts within a domain and the relationship between those
concepts. Ontologies can be based on different assumptions, for specific domains and
different purposes. Thus it is very difficult to compare them using objective metrics.
There are several ontologies built for the English language. For smaller European
languages, one of the most important general ontologies is Princeton Wordnet [8]. It
Shallow Ontology Based on VerbaLex
115
contains many relations (e.g. hypo/hyperonym, is part of) connecting synsets (synonym
set) which are equated with “senses”. Specifically, according to WordNet’s on-line
glossary, a sense is a “a meaning of a word in WordNet. Each sense of a word is in a
different synset”. Princeton WordNet is available under free license also for commercial
applications.
EurowordNet [12] and Balkanet [11] were projects to localize (and improve) parts of
the original version to Central and South East European languages. Thanks to ILI (inter
lingual index), it is possible to connect ontologies and use the result as a multilingual
dictionary. Unfortunately some of the problems of original WordNet still remain [4],
e.g. the assumption that membership in two or more synsets is equivalent to having more
different senses. Some of the WordNet senses are indistinguishable from one another by
any criterion. Attempt to build a WordNet-like ontology for new language was described
in various papers [9], [1]. Creation of proper synsets and assigning the relations is a
time-consuming process that needs expert in this field. One of the most serious problems
of the EWN data is their very strict license.
3
VerbaLex
The basic sentence frame is driven by the lexical characteristics of its predicative constructions based on the set of possible verb valencies of the particular verb. For Czech
languages we posses several lexicons of verb valencies. The first resource is Czech
WordNet valency frames dictionary, which was created during Balkanet project and
contains semantic roles and links to the Czech WordNet. The other resource, Vallex 1.0
[3] is a lexicon based on the formalism of the Functional Generative Description (FGD)
and was developed during Prague Dependency Treebank [2] project. Latest project
named VerbaLex [5] comprehends all the information found in the previous resources
plus adds other relevant ones such as verb aspect, verb synonymy and semantic verb
classes based on VerbNet project [10]. VerbaLex contains 10 478 verbs, 21 123 verb
senses and 19 360 valency frames. Information in VerbaLex is stored in the form of
complex valency frames (CVF).
Complex valency frame is designed as a sequence of elements which form a list of
necessary grammatical features (e.g. preposition and grammatical case).
opustit:4/leave office:1 (give up or retire from a position)
obl
frame: AG <person:1>obl
who1 VERB ACT<job:1>what4
example: Jarek opustil zaměstnání / Jarek left his job
Example sentence can show us that if “Jarek” has to be the agent (semantic role)
then it has to be in nominative (numbered 1) case. Also it has to be a hyponym of
person:1 in the WordNet ontology. Thanks to ILI we can have nodes named in English
and use words from Czech EuroWordNet.
This notation exported to an XML format allows us to easily process both syntactic
and semantic layer of the sentence.
116
4
Marek Grác
Syntactic analyzer synt
The NLP laboratory at FI MU Brno develops a deep syntactic analyzer of Czech sentences, synt [6]. The system uses a meta-grammar formalism which enables it to define
a grammar with maintainable number of meta-rules. These meta-rules are produced
manually by linguists. Meta-rules are translated into context-free rules supplemented
with additional contextual constraints and semantic actions. Efficient and fast headdriven chart parsing algorithm is used for context-free parsing. Final result of parsing is
a set of derivation trees – unfortunately, for more complicated sentences we can obtain
billions of those. After modification of the parser and grammar [7] it is possible to
obtain syntactic structures corresponding to the given nonterminal. With this feature it
is possible to use this tool not only as a deep but also as a shallow parser which identifies
most of important syntactic structures unambiguously (as opposed to the possibly huge
number of derivation trees).
5
Sholva
The Sholva ontology, currently being developed at FI MU, attempts to create a new
lexical resource for Czech language which will be free to use for any purpose. Proposed
methods and implementation of the tools are also suitable for other languages with
similar structure (e.g. Slovak) and we believe that those languages will follow soon.
Our ontology should be used primarily for machine translation, syntactic parsing and
word sense disambiguation. Application driven extensions should be possible.
Our attempt is based on an assumption that creation of a rich lexical resource is
time-consuming and very often the subjective view is one of the most serious problems.
Thus our methodology aims to limit the expert (linguist) as much as possible. We believe that language constructions are obtainable from corpora, so only usages found in
corpora are taken into account.
Shallow Ontology Based on VerbaLex
117
We do not intend to create an ontology with dozens of relations and complicated
structures and we would rather like to follow the KISS methodology (Keep It Simple,
Stupid). For our usage, only hypo/hyperonymical relations can be used directly, and
basic ontology will not contain any other relations. In EWN, senses of words are numbered, but splitting word senses is also a very subjective task. More importantly, this has
no direct relevance to our primary goals. For word sense disambiguation, we need to
be able to distinguish various senses of a given word. We believe that for this purpose,
the knowledge of path from the root node to any given word created by hyponymical
relations defines word sense. It is possible that a word will have several hyperonymical
relations, but we do not know if they refer to the same or a different thing.
We know where we will obtain data for processing and which relations we will
focus on. The process of creating an ontology in this style is very similar to corpora
annotation. Annotation tool will provide everything to the expert who will just select
the correct option or mark an error (syntactic analyzer error or a missing valency frame).
Users will have to select line(s) which contain correct valency and mapping of verb roles
to syntactic structures (mostly preposition and noun phrases). Lines which fill all obligatory slots are marked green, all others are marked red (and cannot be selected). Problems may occur if a slot contains logical OR (e.g. abstraction:1 or communication:2 or
info:1), in which case, we will just assign this formula to the syntactic structure. Later,
when our structure will occur in a more specific slot, we can disambiguate it. It may
also be possible that the given word does not need to be distinguished at all and is only
assigned to a slot within this formula. In that case, the ontology hierarchy and valency
frames should be modified to better correspond to the language usage.
6
Conclusions
In this article about Sholva, Shallow Ontology Based on Valency Frames, we have
presented the basic ideas. Moreover, motivation and reasons for different ontology style
have been explained and a methodology for building such an ontology on top of existing
resources has been given.
In the future there will be further work on annotating large enough corpora – this
should have positive effects on quality of VerbaLex valencies and parsing of syntactic
structures in synt. Existence of such lexical resource suitable for machine processing
should help in our major areas of interest – syntactic parsing and machine translation.
Moreover, after creation of a similar resource for other languages, it will also be possible
to create a multilingual dictionary.
7
Acknowledgements
This work has been partly supported by the Academy of Sciences of Czech Republic
under the project 1ET100300419 and by the Ministry of Education of CR within the
Center of basic research LC536 and in the National Research Programme II project
2C06009.
118
Marek Grác
References
[1] Erjavec, T. and Fišer, D. (2006). Building Slovene WordNet. In 5th Conference
on Language Resources and Evaluation (LREC 2006), Genoa, Italy.
[2] Hajič, J. (1998). Building a syntactically annotated corpus: The Prague dependency treebank. Issues of valency and meaning, pages 106–132.
[3] Hajič, J., Panevová, J., Urešová, Z., Bémová, A., Kolár, V., and Pajas, P. (2003).
PDT-VALLEX: Creating a Largecoverage Valency Lexicon for Treebank Annotation. In Proceedings of The Second Workshop on Treebanks and Linguistic
Theories, volume 9, pages 57–68. Citeseer.
[4] Hanks, P. and Pustejovsky, J. (2005). A pattern dictionary for natural language
processing. Revue française de linguistique appliquée, (2005/2):63–82.
[5] Hlaváčková, D. and Horák, A. (2005). Verbalex–new comprehensive lexicon of
verb valencies for czech. In Proceedings of the Slovko Conference. Citeseer.
[6] Horák, A. and Kadlec, V. (2006). Platform for Full-Syntax Grammar Development
Using Meta-grammar Constructs. In Proceedings of the Paclic.
[7] Jakubíček, M., Horák, A., and Kovář, V. (2009). Mining Phrases from Syntactic
Analysis. In Text, Speech and Dialogue.
[8] Miller, G. (1995). WordNet: a lexical database for English.
[9] Pala, K. and Smrž, P. (2004). Building Czech wordnet. Romanian Journal of
Information Science and Technology, 7(1-2):79–88.
[10] Schuler, K. (2005). VerbNet: A broad-coverage, comprehensive verb lexicon.
Univ. of Pennsylvania-Electronic Dissertations.
[11] Stamou, S., Oflazer, K., Pala, K., Christoudoulakis, D., Cristea, D., Tufis, , D.,
Koeva, S., Totkov, G., Dutoit, D., and Grigoriadou, M. (2002). BALKANET A
Multilingual Semantic Network for the Balkan Languages. In Proceedings of the
International Wordnet Conference, Mysore, India, pages 21–25.
[12] Vossen, P. (1998). Eurowordnet a multilingual database with lexical semantic
networks. Computational Linguistics, 25(4).
Multimodal Russian Corpus (MURCO):
General Structure and User Interface
Elena Grishina
Institute of Russian Language RAS, Moscow, Russia
[email protected]
Abstract. In the paper the Multimodal Russian Corpus (MURCO) is described
in outline. The MURCO is planned to become accessible for general use at the
beginning of 2010. Just now the group of the Corpus developers prepares the
MURCO pilot project. Therefore, the discussion concerning its general structure and interface characteristics is quite urgent.
1 What is MURCO?
The MURCO is being elaborated within the framework of the Russian National Corpus (RNC) and may be considered as the further development of the spoken
component of the RNC. The latter includes the Spoken Subcorpus proper and the prosaic part of the Accentological Subcorpus1. The Spoken Subcorpus is the collections
of transcripts of the spoken texts of different types (private speech, public speech, and
movie speech-tracks) (see [1], [2], [5], [12]). These transcripts are annotated morphologically and semantically according to the RNC annotation system (see [14], [18]
[16]), and also they are annotated from the metatextual point of view (see [17]). In
addition, the Spoken Subcorpus has its own annotation: the accentological annotation
and the sociological one. The accentological annotation presupposes that in every
wordform the real (not normative) stress is marked. Therefore, a user can investigate
the history of Russian accentological system and its normative requirements, which
are specific for this or that period. The sociological annotation means that to every
token the information concerning the sex, the age and the name of a speaker is
assigned, so a user can form his own subcorpora according to all these parameters and
their combinations.
Thus, the Spoken Subcorpus of the RNC (its volume just now is circa 8,5 million
tokens) gives a user the possibility to solve different types of tasks (lexical, morphological, semantic, accentological, sociological and mixed), but all these tasks must not
be connected or based on the real phonation. Therefore, the problems concerning
phonetics, orthoepy, intonation, and also the morphological, semantic, accentological,
sociological problems, which are founded on any aspect of phonetics, have no chance
to be solved by means of the data of the Spoken Subcorpus of the RNC.
1
The poetic part of the Accentological Subcorpus includes Russian poetry of 18–21 centuries
with the marked arses (potentially stressed syllables). It gives a user the possibility to figure
out the real stress of a wordform according to the simple set of the rules (see [6], [10]).
120
Elena Grishina
Such are the reasons for the creation of the MURCO. The main principle of the
Multimodal Russian Corpus is the alignment of the text transcripts and the parallel
sound- and video tracks (see [11], [7], [8]). Consequently, when a user makes his data
query he may obtain not only written text, annotated from different points of view,
but also the corresponding sound- and video material. This possibility let an investigator use the obtained information at his will: he may make use of his own manner of
phonetic transcription, of the elected speech and intonation analyzers; he may pose
and solve all types of research tasks, connected with phonetics, and so on.
Quite a conundrum for the developers of the multimodal corpora is the whole set
of problems concerning copyright offence and privacy invasion (see proceedings of
[15]: every paper, pertinent to the multimodal subject area, contains the debate on the
point). It is the main reason why the multimodal corpora usually are not accessible for
general use. To avoid the range of the copyright and privacy problems, we have
decided to develop the MURCO on the basis of movie subcorpus of the Spoken Subcorpus of the RNC2. There are some features, which distinguish the natural spoken
speech and the cinematographic one (first of all we mean the parameter of the text
coherence), but the differences though remarkable are not crucial (see [3], [4] about
the usage of the discourse markers in the movie transcripts; the strategy of their usage
is virtually the same in the natural and cinematographic spoken Russian); that is to
say, the higher coherence of the movie transcripts in comparison with the transcripts
of the natural spoken texts does not turn the former into the written texts: they remain
spoken ones.
The total volume of the movie transcripts in the RNC is circa 5 million tokens. We
plan to include the movie subcorpus in the MURCO wholly, without reservation. As a
result we mean to obtain the first-rate generally accessible multimodal corpus.
2 MURCO annotation
The types of annotation in the MURCO have been described in our earlier papers (see
[11], [7], [9]), so in this paper we only summarize the theme. Let alone the standard
RNC annotation (see above), the MURCO annotation includes the following items:
•
orthoepic annotation: combinations of sounds are marked
•
annotation of accentological structure: the word structure from the point of
view of stress position is defined
•
speech act annotation: the types of speech acts and vocal gestures, used in a
clip are described
•
gesture annotation: the type of gesticulation in a clip is described
2
Later on we plan to include in the MURCO the patterns of the public spoken speech, which
also does not involve the problems of the kind.
Multimodal Russian Corpus (MURCO)
121
So, the summarized scheme of the MURCO annotation seems to be as follows:
Method of
annotation
Assigned
To text
To word
To clixt
(text+clip)
Automatic
(obligatory)
Semi-automatic
(obligatory)
Manual
(selected)
metatextual
annotation
morphological,
semantic, o r t h o e p i c annotation,
annotation of
accentological word
structure
metatextual
annotation
–
–
sociological,
accentological
annotation
–
–
speech act
and g e s t u r e
annotation
Table 1.
Naturally, the types of annotation, which are specific for the MURCO (in the
Table 1 they are spaced out), define the peculiarities of the MURCO interface in comparison with the interface of the RNC proper. So below we’ll describe only the items
of the MURCO interface, which differ from the RNC interface.
3 Characteristics of MURCO interface
3.1 Orthoepic queries
The orthoepic annotation in the MURCO is founded on the morpho-phonemic principle of the Russian orthography, which means that there are quite transparent
correspondence between the word and word-combination spelling and the word pronunciation. Therefore, we receive the possibility to annotate the combinations of
letters to obtain the pronunciation of the correspondent sounds.
The crucial types of sound combinations in Russian are as follows:
•
C…C = combination of two or more consonants within the word limits
•
V…V = combination of two or more vowels within the word limits
•
C…C#C…C = combination of consonants at the word boundaries
•
V…V#V…V = combination of vowels at the word boundaries
•
C…C#V…V = combination of the consonants before the vowels at the word
boundaries
122
Elena Grishina
Obviously, it is quite easy to annotate such combinations of letters in a text automatically. Consequently, to any tokens in the MURCO will be assigned the set of the
letter combinations, which may present a difficulty. Naturally, all these combinations
suppose to become searchable.
For example, we may formulate our query something like this: *NS* (that is, ‘to
find all tokens, which include the letter combination -ns-), and receive as a result the
clixts, in which may be found tokens like this: pensija ‘pension’, pansion ‘boardindhouse’, Pensil’vanija ‘Pennsylvania’, mansarda ‘attic floor’, tangens ‘tangent’ (and
their derivatives and indirect forms), and so on. If we request *N#S* (that is, ‘to find
all tokens, which have the letter combination - n- and - s- at their boundaries’), we
receive clixts with the combinations of tokens like this: on sejchas ‘he now’, voron sel
‘a raven has sat’, and so on. Obviously, it is very hard to obtain the tokens including
the letter combinations of the kind without special orthoepic annotation. From the
other hand, the queries of the kind may be very useful. For example, the query
described above gives us the possibility to investigate the pronunciation of the letter
combination - ns- before the front vowels. It is well known that in Russian its pronunciation may be of two kinds: [n’s’] with palatalized [n’] and [ns’] with hard [n]. The
orthoepic annotation assists us to define the main factors, which have an influence on
the choice between these two possibilities (the choice depends on the degree of the
assimilation of this or that word, on the age of a speaker and on the presence/absence
of the word boundary between two consonants).
3.2 Queries on accentological word structure
It is well known that the dynamic quality of Russian stress leads to the great degree of
the reduction of the unstressed syllables in a word. Consequently, it is very important
to give a user an opportunity to obtain information, concerning the position of the
stressed syllable and the quality of the stressed vowel, the position/quality of the preand post-tonic vowels, and so on. Owing to the fact that the majority of the clixts in
the MURCO are accentuated, it is possible to annotate the accentological structure of
any token in automatic mode. The content of the possible requests is defined in line
with the Table 2:
quality of vowel
number of syllable
stressed vowel
A
1
pre-tonic vowel
B
2
post-tonic vowel
C
3
quantity
syllables
4
Table 2.
Multimodal Russian Corpus (MURCO)
123
In the table cells A–C a user may specify the letter designation of a vowel (in the
stressed, pre- and post-tonic syllable), in the cells 1–3 – the number of the corresponding syllable, in the cell 4 – the quantity of the syllables in a word. All these
parameters are independent, so a user can freely combine them if necessary. For
example, a user may request all tokens containing 1) the second post-tonic syllable, 2)
the stressed syllable o, 3) three syllables, 4) vowel o in the second pre-tonic syllable,
while a token has 4 syllables and the stressed vowel o.
All these parameters are very important for the phoneticians, specialists in orthoepy, dialectologists, and investigators in the area of the history of Russian. In
addition, the importance of orthoepic and accentological annotation can scarcely be
overestimated, having in mind the professional interests of the teachers of Russian,
uppermost as a foreign language.
3.3 Speech act queries
The specialists on Russian intonation, linguistic pragmatics, socio- and psycholinguistics will obtain using the MURCO very significant information on different
aspects of the speech acts. The acquisition of the information of the kind is provided
with the special annotation, which is attached to a clixt in manual mode. Earlier we
have described the basic principles of this annotation (see [7], [8]) and the software to
hasten the process of the annotation and to unify its results (see [13], [9]).
In this paper we would like to present the items of the MURCO interface, connected with this type of annotation.
3.3.1 Sociolinguistic characteristics of clixt
• Quantity of participants (1, 2, 3, many)
• Participants’ sex (Mas, Fem, Mixed)
• Language (Russian, Russian with accent, Foreign (Ukrainian, English, and so on),
Quasi, Secret… the list is open)
• Social situation (Telephone call, Dinner speech, Talk with authorities… the list is
open)
3.3.2 Types of speech acts in clixt
1.
2.
3.
4.
5.
6.
7.
Address or call
Agreement
Assertion
Citation
Complimentary
Critical utterance
Etiquette formula
8. Imperative
9. Joke
10. Modal utterance or performative
11. Negation
12. Question
13. Trade utterance
124
Elena Grishina
Naturally, every type has the set of its own subtypes, for example,
•
Question
− Closed
− Leading
− Open
− Indirect
− Inarticulate
− Contact
− Overinterrogation
− Incomprehension
− Feedback
− Critical
− Question to oneself
− Etiquette
•
Joke
− Irony
− Sneer
− Hint
− Joke
Till the moment we pick out 150 subtypes of different speech acts grouped into 13
types, mentioned above. The majority of the speech act types are based on the Russian verbs, describing the process of speech. Consequently, they are intuitively
comprehensible by all Russian speakers. To every clixt may be attached more than
one type of speech act, and moreover, every speech act in a clixt may be characterized
from different points of view (i.e., an assertion may be characterized at the same time
as information, declaration, statement, and some others). Thus, the classification of
speech act is not tree-like, it is faceted.
3.3.3 Completeness of utterances
•
•
•
•
•
•
•
•
Full
Self-interruption
Question without answer
Overlapping cues
Unfinished
Interrupted
Continued
Gesture instead of word
3.3.4 Repetitions and their types
−
−
−
−
−
−
•
No repetitions
•
Types of repetitions
Single
Multiple
One-word
Many-word
Overinterrogation
Repetition with intensifier
−
−
−
−
−
−
Repetition with different intonation
Echo
Relay
Mimicking
Repetition with the change of listener
Envelope repetition
Multimodal Russian Corpus (MURCO)
3.3.5 Manner of speech
•
•
•
•
•
•
•
•
•
•
•
Normal
All together
Articulation disorders
Chanting/scanning
Crying
Declamation
Dictation
Drunken
Dubbing-in
Fast
Humming
•
•
•
•
•
•
•
•
•
•
Inarticulate
Laughing
Parceling out
Reading
Shout
Slip of the tongue
Talking to oneself
Ventriloquism
Voice-over
Whisper
−
−
−
−
Hemming
Iconic sounds
Stopgaps
Teasing sounds
3.3.6 Vocal gestures and interjections
•
−
−
−
−
•
−
−
−
−
−
−
−
−
•
−
−
−
−
−
−
−
−
Vocal gestures
Address to an animal
Beep
Click of the tongue
Feeling cold
Interjections
− Admiration
− Agitation
− Agreement
− Approval
− Attraction of smb’s attention
− Backing-hum
− Backing-well
Backing-yes
Backing-вот
Bewilderment
Chagrin
Comprehension
Dissatisfaction
Distrust
Fright
Physiological activities
Breathe the air
Chuckle
Cough
Exercise stress
Exhalation
Groaning
Kiss
Sigh
−
−
−
−
−
−
−
−
− Grief
− Indignation
− Mockery
− Negation
− Pain
− Pity
− Question
Recollection
Reply to address
Scorn
Sudden recollection
Surprise
Threat
Unexpectedness
Whoops!
−
−
−
−
−
−
−
Smacking one’s lips
Sniffing
Spit
Taking one’s breath
Weeping
Whistle
Yawn
125
126
Elena Grishina
These are the main point of the MURCO interface, concerning the speech act
annotation. As follows from the written above, the MURCO let a user collect very
diverse material as for the speech activities. This fact opens up new possibilities for
the investigators of spoken Russian.
3.4 Gesture queries
During last three decades the investigation of the role of the gesticulation in different languages has progressed to a large degree. Now it is the current opinion that it is time to
elaborate the gesture corpora to base the investigation of the gesture systems on a hard
ground (see the materials of [15] and their review and the main bibliography in [7]).
The MURCO seems to be the resource, which is generally accessible and quite
considerable as for its volume, moreover, the MURCO is planned to include a lot of
video tracks. So, it is absolutely necessary to provide a user with the annotation and
interface concerning Russian gesticulation.
The basic principles and ideological grounds for our gesture classification we gave
described earlier (see [7]). So, in this paper we list the main items of the MURCO
interface, concerning the gesticulation subject matter.
3.4.1 Sociological characteristics of gesture
•
Name of a speaker (if known)
•
Speaker’s sex (Mas, Fem, Unknown)
•
Character’s sex (Mas, Fem, Unknown, Man, playing woman, Man, pretending to be woman, Woman, playing man, Woman, pretending to be man)
•
Speaker’s age (baby, teenager, adult, elderly, unknown)
•
Character’s age (baby, teenager, adult, elderly, unknown)
3.4.2 Repetition factor
•
•
Single gesture
Multiple gesture
3.4.3 Active organ
•
−
−
−
−
−
−
−
−
Main organ: head
brow
brows
chin
ear
eye
eyes
face
forehead
−
−
−
−
−
−
−
−
head
lips
lower lip
mouth
nose
tongue
upper lip
upper teeth
Multimodal Russian Corpus (MURCO)
•
−
−
•
−
−
−
−
Main organ: body
body
shoulder
Main organ: arm
arm
fingers
forefinger
forefinger+long finger
−
forefinger+long finger+fourth finger
−
•
−
−
•
−
•
−
forefinger+long finger+thumb
Main organ: arms
arms
hands
Main organ: leg
foot
Main organ: legs
feet
−
−
shoulders
back
−
−
−
−
−
−
forefinger+thumb
fourth finger
hand
little finger
long finger
thumb
−
−
forefingers
fingers
−
shin
−
legs
127
3.4.4 Passive organ
The set of the passive organs is specific for this or that active organ. The basic
passive organs are as follows:
−
No passive organ
−
hand
−
arm
−
head
−
arms
−
hip
−
back
−
hips
−
body
−
lips
−
breast/stomach
−
lower lip
−
chin
−
mouth
−
eat
−
neck
−
eye
−
nose
−
face
−
shoulder
−
fingers
−
throat
−
hair
3.4.5 Adaptor
Adaptor is the object, which is a necessary component of this or that gesture, but
is not one of the human organs. The main types of adaptors are as follows:
−
No adaptor
−
external object
−
cloth
−
glasses
−
earth
−
gloves
128
Elena Grishina
−
−
−
−
−
−
−
handset
headwear
heavy object
interlocutor
piece of furniture
pocket
sky
−
−
−
−
−
−
surface
tableware
tie
vessels
watch
wristlet
3.4.6 Dimensional characteristics of gesture
•
−
−
−
•
−
−
−
−
−
−
−
−
Palm orientation
up
down
one opposite the other
Direction of movement
backward
differently directed
does not matter
downwards
forward
forward-backward
from right to left
from the outside to the center
−
−
−
to speaker’s body
outside
perpendicularly to speaker’s body
−
−
−
−
−
−
−
−
from within outside
horizontal circle
on its axis
outside
to oneself
to the center
upwards
vertical circle
3.4.7 Gesture meanings and gesture types
Till the moment we have marked out about 250 gesture meanings, which are grouped
into 14 gesture types. The gesture types are as follows (see at greater length in [8]):
•
Adopted
•
Gestures – speech acts
•
Conventional
•
Gestures of inner state
•
Corporate
•
Iconic
•
Critical
•
Physiological
•
Decorative
•
Regulating
•
Deictic
•
Rhetorical
•
Etiquette
•
Searching
Every type includes some gesture meaning. For example, some of the etiquette
gestures are as follows:
• gratitude (to applaud, to move one’s head forward, twice-repeated kiss, to close
one’s eyes, to nod, to touch smb, to bow, to touch smb’s hand, to kiss smb, to kiss
smb’s hand, press one’s hands to one’s breast, and so on)
Multimodal Russian Corpus (MURCO)
129
• apology (to beat one’s breast, to nod, to move one’s chin outside, to press smb’s
hand to one’s breast, to press one’s hand to one’s breast)
• invitation (to nod, to show smth with one’s hand, to bow), and so on.
4 Conclusions
Thus we can see that the MURCO considerably extends searching possibilities up
about the characteristics of spoken Russian. We may illustrate the fact with the queries, applying to the Russian greeting formulas (GF).
Corpus
Types of queries
(3) Lexical queries: the retrieve of the
specific lexemes, used in GF (i.e.
zdravstvujte ‘how do you do?’, privet
‘hi!’, and so on)
(4) Morphological queries: the retrieve
of the specific morphological characteristics of the GF lexemes (i.e.
zdravstvujte (Pl or courtesy) vs
zdravstvuj (Sg), privet (Noun) vs
privetstvuju (Verb), and so on)
(5) Sociological queries: the forming of
the gender and chronological subcorpora to investigate the peculiarities
of the GP usage
(6) Semantic & speech act queries: the
retrieve of all Russian GP simultaneously
(7) Orthoepic/accentological queries:
the retrieve of the types of the vowel
contractions and the shortening of
the consonant groups in the GF; the
reduction of the pre- and post-tonic
vowels in GF
(8) Speech act queries: the retrieve of
the types of repetitions, used in GF;
the types of vocal gestures and interjections, accompanying the different
types of GF; GF, used in the
man/woman dialogues; see also the
item 4
(9) Gesture queries: the retrieve of the
gestures, accompanying Russian GF
Spoken Subcorpus of
RNC
+
MURCO
+
+
+
+
–
+
–
+
–
+
–
+
Table 3.
+
130
Elena Grishina
Acknowledgement
The work of the MURCO group is supported by the program “Genesis and Interaction
of Social, Cultural and Language Communities” of the Russian Academy of Sciences.
The author’s investigation is supported by the RFBR 3 (RFFI) under the grant 08-0600371а.
References
[1] Grishina, E. (2005). Ustnaja rech v Nacional’nom korpuse russkogo jazyka.
Nacional’nyj korpus russkogo jazyka: Rezul’taty i perspektivy. M.: 94–110,
available at: http://docs.google.com/Doc?id=df52fjjj_12fcbpqbdq&hl=en
[2] Grishina, E. (2006). Spoken Russian in the Russian National Corpus (RNC).
LREC2006: 5th International Conference on Language Resources and Evaluation. ELRA: 121–124, available at:
http://docs.google.com/Doc?id=df52fjjj_3wd9mcrdg&hl=en
[3] Grishina, E. (2007). O markerah razgovornoj rechi (predvaritel’noje
issledovanije podkorpusa kino v Nacional’nom korpuse russkogo jazyka).
Kompjuternaja lingvistika i intellektual’nyje tehnologii. Trudy mezhdunarodnoj konferencii “Dialog 2007”. M.: 147-156, available at:
http://docs.google.com/Doc?docid=df52fjjj_0d5gbwscn&hl=en,
http://www.dialog-21.ru/dialog2007/materials/html/22.htm
[4] Grishina, E. (2007). Text Navigators in Spoken Russian. Proceedings of the
workshop “Representation of Semantic Structure of Spoken Speech”
(CAEPIA’2007, Spain, 2007, 12-16.11.07, Salamanca). Salamanca: 39-50,
available at: http://docs.google.com/Doc?docid=df52fjjj_11fmxsdzdh&hl=en
[5] Grishina, E. (2008). Ustnaja rech v Nacional’nom korpuse russkogo jazyka.
Sbornik trudov. XX sessija Rossijskogo akusticheskogo obsh’estva. M.: 88–91,
available at: http://docs.google.com/Doc?docid=df52fjjj_2gf3httgq&hl=en
[6] Grishina, E. (2009). Korpus “Istorija russkogo udarenija”. Nacional’nyj korpus
russkogo jazyka: 2006–2008. Novyje rezul’taty i perspektivy. M.-SPb.: 150–
213
[7] Grishina, E. (2009). Mul’timedijnyj Russkij Korpus (MURCO): problemy
annotacii. Nacional’nyj korpus russkogo jazyka: 2006–2008. Novyje rezul’taty
i perspektivy. M.-SPb. (forthcoming)
[8] Grishina, E. (2009). Nacional’nyj korpus russkogo jazyka kak istochnik
svedenij o russkoj rechi. Rechevyje tehnologii. (3). M. (forthcoming)
[9] Grishina, E. Multimodal Russian Corpus (MURCO): types of annotation and
annotator’s workbenches. Proceedings of Corpus Linguistic Conference 2009
(UK, Liverpool, 20–23 July). Liverpool (forthcoming)
3
The Russian Fund of Basic Researches.
Multimodal Russian Corpus (MURCO)
131
[10] Grishina, E., Korchagin, K., Plungian, V., Sichinava, D. (2009). Poeticheskij
korpus v ramkah Nacional’nogo korpusa russkogo jazyka: obsh’aja struktura i
perspektivy ispol’zovanija. Nacional’nyj korpus russkogo jazyka: 2006–2008.
Novyje rezul’taty i perspektivy. M.-SPb.: 71-113
[11] Grishina, E, Savchuk, S. (2008). Korpus zvuchash’ej russkoj rechi v sostave
Nacional’nogo korpusa russkogo jazyka. Proekt. Kompjuternaja lingvistika i
intellektual’nyje tehnologii. Trudy mezhdunarodnoj konferencii “Dialog
2008”. M.: 125-132, available at:
http://docs.google.com/Doc?docid=df52fjjj_1dzj8gchs&hl=en,
http://www.dialog-21.ru/dialog2008/materials/html/19.htm
[12] Grishina, E., Savchuk, S. (2009). Ustnyj korpus v Nacional’nom korpuse
russkogo jazyka: sostav i struktura. Nacional’nyj korpus russkogo jazyka:
2006–2008. Novyje rezul’taty i perspektivy. M.-SPb (forthcoming)
[13] Kudinov, M., Grishina, E. (2009). Instrumenty poluavtomaticheskoj razmetki
dl’a Mul’timedijnogo russkogo korpusa (MURCO). Kompjuternaja lingvistika
i intellektual’nyje tehnologii. Trudy mezhdunarodnoj konferencii “Dialog
2009”. M.: 248–261, available at:
http://www.dialog-21.ru/dialog2009/materials/html/40.htm
[14] Kustova, G., Lyashevskaja, O., Paducheva, E., Rahilina, E. (2005).
Semanticheskaja razmetka leksiki v Nacional’nom korpuse russkogo jazyka.
Nacional’nyj korpus russkogo jazyka: Rezul’taty i perspektivy. M.: 155–174,
available at:
http://ruscorpora.ru/sbornik2005/10kustova.pdf
[15] LREC (2008). 6th International Conference on Language Resources and
Evaluation. Marrakesh: ELRA. Available at:
http://www.lrec-conf.org/proceedings/lrec2008/
[16] Lyashevskaja, O., Plungian, V., Sichinva, D. (2005) O morfologicheskom
standarte Nacional’nogo korpusa russkogo jazyka. Nacional’nyj korpus
russkogo jazyka: Rezul’taty i perspektivy. M.: 111-135, available at:
http://ruscorpora.ru/sbornik2005/08lashevs.pdf
[17] Savchuk, S. Metatekstovaja razmetka v Nacional’nom korpuse russkogo
jazyka: bazovyje principy i osnovnyje funkcii. Nacional’nyj korpus russkogo
jazyka: Rezul’taty i perspektivy. M.: 62-88, available at:
http://ruscorpora.ru/sbornik2005/05savchuk.pdf
[18] Sichinava, D. (2005). Obrabotka tekstov s grammaticheskoj razmetkoj:
instrukcija razmetchika. Nacional’nyj korpus russkogo jazyka: Rezul’taty i
perspektivy. M.: 136-154, available at:
http://ruscorpora.ru/sbornik2005/09sitch.pdf
Electronic Lexical Card Index for the Ukrainian
Dialects (ELCIUD)
Pavlo Grytsenko, Olena Siruk, and Viktor M. Sorokin
Institute for the Ukrainian Language, Ukrainian Academy of Sciences
Laboratory for Computational Linguistics,
National Taras Shevchenko University of Kyiv, Ukraine
[email protected]
Abstract. The aim of this paper is to present an Electronic Lexical Card Index
for the Ukrainian Dialects (ELCIUD). Our investigation focuses on the main
problems that occur during compilation process (searching of the optimal structure of a database and an electronic dialect card, granting a possibility of editing
all the linguistic information inside and adding a new one taking into account
differences in localization marking systems of different authors etc.) and possible
ways of resolving them. We also review future trends of ELCIUD development,
especially its transformation into a computational lexicographical system, and
using it as a base for further linguistic experiments as computer means of dialect
text research.
Keywords: dialect, lexical databases, computational lexicographical systems.
1
Topicality of the research
The computer-aided design of lexicographic systems satisfying the demands of today’s
information-aware society for the systematized and freely accessible linguistic data is
an important task of modern Ukrainian lexicography. The goal is to get comprehensive
and detailed information quickly into a form suitable for its processing in accordance
with the formulated linguistic task. A computer dictionary, providing at least the same
degree of thoroughness of a conventional paper dictionary, enables us to extend substantially the information scope of the lexicographic system, to better prepare language
material for subsequent processing and in general to increase the speed of work. Consequently, the design of computer dictionaries, card indexes and additional software is
of pressing concern.
It is a question of great importance not only for literary idiom, but also for dialects.
In recent years dialectology has taken up the use of electronic sources of information (such as electronic dialectal atlases, text corpora of interviews with respondents,
postal questionnaires and historical text corpora) and electronic methods (such as computer mapping, statistical processing of data, collation with atlases and presentation on
maps)1 . Thus attention is concentrated on dialectometry, defined as the mathematically
1
The first attempts at the application of computer methods for language data processing were made by Slovak researchers, in particular P. Žigo for the “Slavic Linguistic
Atlas”. In the 1990s Žigo promoted the advantages of working with computers and
Electronic Lexical Card Index for the Ukrainian Dialects
133
expressed degree of identity∼difference between the compared dialectal systems (with
the employment of frequency analysis and orientation for psychological conclusions)2
or on the closeness∼remoteness of the phenomena in compared dialects or dialectal
languages [8]. It is self-evident that the application of mathematical factors during the
comprehension of dialectal materials and the use of computer technologies brings with
it increased perspectiveness and expedience. Similarly, the use of computer technologies for database preparation, modeling, and development of dialectal dictionaries as
information generators with respect to a dialectal language brings great benefits [1].
The experience of computational lexicography [14], based largely on processing of the
the literary or standard language, can also be of great help in state-of-the-art dialectal
lexicography.
2
Theoretical principles of the project
Electronic Lexical Card Index for the Ukrainian Dialects (ELCIUD) is a linguistic
database (LDB), created on the basis of published dialect dictionaries and unpublished
detailed his experience of collecting dialectal material in Slovakia [2]. Polish scholars
began to use statistical methods for dialectological research at around the same time
[10; 15]. Today in Poland there is an online information retrieval system dealing with
dialects and dialectology: the project “Gwary polskie: Przewodnik multimedialny”,
edited by H. Karaś, was completed in 2009 [9]. The informatization of linguistic
research is developing actively in the Russian Federation. In particular, the project
“Lexical Atlas of Russian National Dialects” is characterized by a research methodology involving computer mapping and a systematic approach to the interpretation
and cartographic presentation of the material [5; 6]. At the beginning of the 1990s,
dialectal texts were seen as one of the basic resources of dialectology: dialects are
investigated as separate communicative systems. This direction was referred to as
“communicative dialectology”. The concept of dialectal corpus creation as an optimal
base for the maintainance and subsequent processing of dialect text material was
utilised, as a dialectal subcorpus, in the composition of the National Russian Corpus
of the Institute for the Russian Language (Russian Academy of Sciences) and the
Saratov Dialectal Corpus (Saratov State University) [3; 4]. The understanding of
dialectology in accordance with this approach can be traced all the way back to
Aristotle, and, with the application of exact calculation methods and the latest
technologies, has reached its highest development in the USA [11; 12; 13].
2 The mathematical modelling of the areal variation of dialectal features in combination with a revision of the traditional understanding of a dialect, according to
William Kretzschmar (“Quantitative Areal Analysis of Dialect Features”), has all
the chances of updating our knowledge about language and its regional differences.
The article is devoted to the quantitative analysis of separate dialectal features from
the “Linguistic Atlas of the Middle and South Atlantic States”. The XII International conference Methods XII (August 2005, Université de Moncton, Canada) was
devoted to the problems of methods in dialectology, to innovative approaches and,
in particular, contemporary theoretical and methodological trends in this area [16].
Within the conference special seminars were held by David Heap (University of
Western Ontario), John Nerbonne (University of Groningen), William Kretzschmar
(University of Georgia).
134
Pavlo Grytsenko, Olena Siruk, and Viktor M. Sorokin
handwritten lexical card indexes of other projects (Ukrainian materials in the lexical
part of the “Slavic Linguistic Atlas”, the “Atlas of Ukrainian” and the “Lexical Atlas of
Ukrainian”).
The goal of the ELCIUD is to gather from different sources all lexical units of
the Ukrainian dialects, to process these as elements of the information system and to
provide operative access to this source of linguistic data. The task of the project carried
out by the Institute for the Ukrainian Language of the Ukrainian Academy of Sciences
and the National Taras Shevchenko University of Kyiv is the creation of the linguistic
and computational complex ELCIUD, which allows formalization of the language
material, and presentation of information (dialectal units) in a structured output, suitable
for computer-aided processing.
3
Stages of work
The realization of the project is divided into three stages. In the first stage, the creation
of a linguistic database of register units is to be made on the basis of dictionary material
(included in the register are dialectal units of all parts of speech). This is the stage of
conversion of present electronic versions of dictionaries into a database which includes:
1) title word; 2) the content part of the vocabulary entry. In the second phase we enlarge
the database by the import of new dictionaries and separate dialectal units, watching out
for and correcting conversion errors, and also correcting the text of vocabulary entries,
without changing their content. At the third stage, formulation of the database structure
is planned by the selection of different types of linguistic information. This is the stage
of formulation of the explanatory part of vocabulary entries and their enrichment by
additional linguistic information.
4
Issues with ELCIUD design
There are several issues which arise with the creation of ELCIUD and whose solution
is of pressing importance. These can be presented in the following key questions:
1.
2.
3.
4.
5.
what kind of linguistic information should be present in the database
how to present this information
how to add new information
what are the possible directions of further data processing
what format should the final product have.
Within regard to the first question, we decided to focus on the vocabulary entries of published dictionaries. However, this does not eliminate the possibility of also adding vocabulary entries from non-printed sources. The second problem required additional experimental research. Working with dialectal dictionaries revealed divergences in structure, filling and presentation of vocabulary entries. This occurred not only with different
authors, but also for the same author [Arkushyn, Grytsak, Magrytska, Onyshkevych,
SUSG]. For example, the structural elements of a vocabulary entry for a monosemantic
noun in the dictionary of Polissya’s dialects [Arkushyn] have clear enough formal signs
and sequence of description (see Table 1).
Electronic Lexical Card Index for the Ukrainian Dialects
 Structural
element
1 Headword
2
3
4
5
6
7
8
Formal signs
Word. Comes after an
indentation. Given in
capital letters.
Phonetic
A word written in special
transcription
symbols is in square
brackets.
Gen. sing. form (pl. Partial word. Begins with a
– for Pluralia
hyphen, ends with a
tantum)
comma.
Marker of gender Italicized lowercase
for a noun or a
letter(s); ends with a point.
number (for
Pluralia tantum)
Stylistic marker
Italicized lowercase
letter(s); ends with a point.
Definition of lexeme Sentence; begins with the
character ‘, ends with the
character ’ and has a point
after it.
Example of use
Text, written in a standard
font, not italic, not bold,
with the additional signs of
phonetic transcription.
Localization of
Number from 1 to 601, in
dialecticism
accordance with the list of
localities produced by the
author, or all of the
populated localities.
Example 1
135
Example 2
ÎÁËIÏËßÍÅÖÜ ÎÁÈÒÀ™ÌÖÈ
[îáë'ïë'àíåö']
[îáèò
àéåìöè]
-íö'à,
-èâ,
÷.
ìí.
êóëií.
absence
ïèðiæîê ç òåðòî¨ çâiði, ÿêi
êàðòîïëi
ïîñòiéíî æèâóòü
íà îäíîìó ìiñöi
absence
463
Òî í
àøè
îáèò
àéåìöè,
æèâ
óò â í
àøîìó
ë'ñè
244
Table 1. Correspondence between structural elements and their formal signs in a
vocabulary entry for a monosemantic noun in [Arkushyn]
As we can see, some cells are irregularly filled (the text-example of the use of
dialecticism is absent in the first example, also as a stylistic marker in the second
example), and this irregularity is pervasive. When it concerns non-optional elements
(e.g. stylistic marker), and basic elements (such as an example of the use of lexeme),
and also when we have extremely syncretic representation of data in a vocabulary entry,
it is not easy both to expand such article to its complete structure and automatize this
process. The structure of the article within the limits of the similar part of speech can
vary substantially, in particular depending on the number of values and pronunciations
of the title lexeme. For example, vocabulary entries for the headword “baba”:
136
Pavlo Grytsenko, Olena Siruk, and Viktor M. Sorokin
ÁÀÁÀ1 [á
àáà] -è, æ. 1. `æiíêà, ÿêà ì๠âíóêiâ' (âñi í. ïï.); 2. `áóäü-ÿêà ñòàðà
æiíêà' (âñi í. ïï.); 3. `äðóæèíà'. À òâîé
à á
àáà âä
îìà? Âîí
à íè áîéö':à
òàêîãî ÷îëî²éêà, éàê ò
è 392, 11, 90, 342; 4. çíåâ. `äiâ÷èíà'. Ê'iðîâí'÷êà
õ
î÷å íàñ â òðåò'îìó ïîñàä
èòè ç áàá
àìè. 392.
ÁÀÁÀ2 [á
àáà] -è, æ. `òñ, ùî ÒÐIÑÊÀËÎ' 282;
ÁÀÁÀ3 [á
àáà] -è, æ. êóëií. `òåðòà êàðòîïëÿ, ðîçëèòà íà äåêî i ñïå÷åíà â
ïå÷i'. Á
àáó òðàáà ñïå÷
è. 496 [Arkushyn].
We can show BABA1 as a tree in accordance with the structure of vocabulary entry for
a noun (with a headword as a root) and will see how many structural elements do not
have their formal representations, either due to absence in general or through the “data
compaction” of the articles in a paper dictionary:
ÁÀÁÀ1
(Phonetic transcription) [áàáà]
(Gen. sing. form) -è
(Marker of gender for a noun) æ.
(Denition 1) æiíêà, ÿêà ì๠âíóêiâ [a woman, who has grandchi-
ldren]
(Stylistic marker 1) N/A
(Example 1) N/A
(Localization 1) (âñi í. ïï.) [all of the populated localities]
(Denition 2) áóäü-ÿêà ñòàðà æiíêà [any old woman]
(Stylistic marker 2) N/A
(Example 2) N/A
(Localization 2) (âñi í. ïï.) [all of the populated localities]
(Denition 3) äðóæèíà [a wife]
(Stylistic marker 3) N/A
(Example 3) À òâîéà áàáà âäîìà? Âîíà íè áîéö':à òàêîãî ÷îëî-
²éêà, éàê òè? [Is your wife at home? Isn't she afraid of such a husband as
you?]
(Localization 3) 392, 11, 90, 342
(Denition 4) äiâ÷èíà [girl]
(Stylistic marker 4) çíåâ. [Disparagingly]
(Example 4) Ê'iðîâí'÷êà õî÷å íàñ â òðåò'îìó ïîñàäèòè ç áàáàìè
[Teacher wants us to sit with girls in the third form.]
(Localization 4) 392
It is evident from the dictionary entry that the example of the use of dialecticism appears
only in entries 3 and 4 of BABA1 . We do not have it for BABA2 . A stylistic marker is
present only in one of four values of BABA1 and in BABA3 .
A tree structure can be much more difficult to produce in cases of several variants
of the phonetic transcription:
ÎÁËÓÄ [Àðêóøèí]
(Phonetic transcription 1) [îáëóä]
(Gen. sing. form) -ó
(Marker of gender for a noun) ÷.
(Denition 1) òñ, ùî ÁËÓÄ [the same as straying/wandering]
(Stylistic marker 1) N/A
Electronic Lexical Card Index for the Ukrainian Dialects
137
(Example 1) Â ë'ñè îáëóä áåðå [Oblud/Straying takes somebody in
the forest. ∼ It is easy to lose one's way in the forest.]
(Localization 1) 129, 111
(Phonetic transcription 2) [îáëóä]
(Gen. sing. form) N/A
(Marker of gender for a noun) N/A
(Denition 1) N/A
(Stylistic marker 1) N/A
(Example 1) Òóò îáëóä âîç'ìå ë'óäèíó [Here oblud/wandering will
take a man.]
(Localization 1) 475, 432
The structure of the article varies according to the part of speech, the number of values,
and the manner of pronunciation of the title lexeme, and accordingly has changing
formal features. Formal features, in turn, can mutate both in accordance with structural
changes and as a result of a change in the formal characteristics of other structural
elements. In particular, the number of headword constituents influences the formal signs
of such structural elements as phonetic transcription and Gen. sing. form (Gen. pl. – for
Pluralia tantum). For example, the headword BAGNOVY LYS [bog fox] consists of two
words, consequently, it has two phonetically transcribed words [áàãíîâ
èé ëèñ] and two
parts of words for Gen. sing. form marking, which begin with a hyphen and end with
one comma – -
îãî -à (ÁÀÃÍÎÂÈÉ ËÈÑ [áàãíîâ
èé ëèñ] -
îãî -à, ÷. `ëèñ, ÿêèé
æèâå íà áîëîòi (â áàãíàõ); ó íüîãî øåðñòü òåìíî-êîðè÷íåâîãî êîëüîðó' 122
[a fox that lives in a bog; its wool is dark-brown]) [Arkushin]. From a linguistic point of
view this phenomenon is simply a linguistic fact, but at the level of the algorithmization
of the addition of language data, this and other features cause additional steps in the
working process.
When the researcher aims at processing many dictionaries in one chart, it is worth
checking in advance how significant the difference in the structure of these dictionaries
is. Compare the vocabulary entries for BABA in four sources (here presented in a single
font face and size):
1. ÁÀÁÀ1 ðîä. îäí. û, ðîä. ìí. ó, óâ, èé, áàá. Ìàòè áàòüêà àáî ìàòåði. Ìîéå
áàáà i éîãî áàáà ïóä éåí:èì ñîíö'îì ïëàò'à ñóøèëè. Ìä. Òàêà ó íàñ ÷åñíà
áàáà, òàê íàñ ë'óáèò. Òí. Ëï÷., Âí, Íâö, Äðâ, Ââð, Áëê, Çëò, ÂÁ÷,ßñ,
Îí, Âðö, ×í. Ïåñòë. áàáóíà, áàáóíêà. Äìø. 2. Ñòàðà æiíêà. Ä'iäîâû áåç
áàáû íèìà ïîðàäû. Ïð÷. Òà êîëèñ' i 0 áàáà áû0 ëà 0 ä'iâêîâ. Ìä. Ãóí'i ñóò'
ëåì õè0 áà ó ñòà0 ðûõ áà0 áó. Êîëî áàá ñòî0 éàëè è ìîë
îä'i 0 æîíè. Êé. Çáð, Ïð, Ñê, Êé, Ëêö, Âí, Êì, Êìí, Äíë, Òõ, Äëâ, Äâ, Ç÷, Ìä, Òðâ.
Çãðóá. áàáåðà. 3. Æiíêà âçàãàëi, äðóæèíà. Éà èç ñ'âîéèâ 0 áàáîâ ó0 ÷îðà
ö'iëèé 0 âè÷'ið áûâ. Òðí. Ìàéå âóí ïðåìíîãî áà0 áèé. Òðâ. Ìîéà áàáà ø÷å
ëèø ñ'iì0 íàö:iò ãîä'i² ìàéå. Òí, Ëï÷, Âí, Òðñ, ËÏë. 4. Çíåâàæë. (ó
ïîðâíÿííi). ×0 êîäà, æå òà0 êó 0 ôàéíó çàðï0 ëàòó 0 ìàéå, à òà0 êà 0 õîäèò',
éàê 0 áàáà. 5. (Çíåâàæëèâî ïðî ÷îëîâiêà). Òî0 òî òâ'i ÷îëîâ'iê íå ÷îëîâ'iê,
àëå 0 áàáà, êî0 ëè òî0 á'i äîçâî0 ë'àéå 0 âóëèö'àìè ñ'à âà0 ë'àòè. Âí, Òðñ,
Ëêö, Âð÷. 7. Ïîâèòóõà. Éó0 ðèøí'ié 0 Ìàð'i ëè0 øå ñ'ië'ñ'0 êà 0 áàáà ïîìîã0 ëà
óðî0 äèòè ä'i0 òèíó. Içê. Íàøà 0 áàáà à0 ëóìñêà 0 äóæå ç0 íàéå 0 êîëî æóí
138
Pavlo Grytsenko, Olena Siruk, and Viktor M. Sorokin
õî0 äèòè. Âí. Ó íàø'iì ñèë'i 0 äîáðà 0 áàáà, áî çà íèâ 0 ä'iòè íè âìè0 ðàâóò'.
Êâ. Içê, Íâö, Áäâ, Êë÷, ×í, Ëâ, Êâ, Äâ, Äìø, Âí. ÀËÀË'Ñ'ÊÀ
(âà0 ëàë'ñ'êà) 0 áàáà, à0 ëóìñ'êà 0 áàáà, âåð0 ìåö'êà 0 áàáà, ñ'ië'ñ'êà
0 áàáà, 0 áàáà-ïóïî0 ð'içêà.
Éàê óâ áóäå îé 0 âàðìåö'êà 0 áàáà ïîìà0 ãàòè, òà
óâ
ä'iò'
àòêî ðåâàòè. Êðö. Éàê æî0 íà õî0 ò'iëà çë'å0 ÷è,
0
0
çàêëèêàëè âà ë'àñ'êó áàáó. Çá. Íâö, Ëâ. 8. Åòí. Óÿâíèé îáðàç çëî¨
ïîòâîðíî¨ ÷àêëóíêè (ó ñïîëó÷åííi ç ÿãà, áîñîðêàíà). Êîé ä'iò'îì êà0 çàòè
0 êàçêó, òà âñå áîéàòñ'à ëèø áàáè áîñîðêàíi. Âí. Ïóä ìîñòîì éå 0 áàáàéà0 ãà. Ãòí. Äâã. 9. åòí. Ñòðàøíà, ïîòâîðíà ñòàðà æiíêà â íàðîäíîìó
ïîâið'¨. Õòî ïåðøèé ðàç äåñ' éäå, äå ø÷å íå áóâ, òà 0 êàæó÷, øî 0 ìàéå
íà äî0 ðîç'i ö'óë'óâàòè ñòðàø0 íó, øìàð0 êàâó òà èìáàòó 0 áàáó à äiâ0 êè
à0 áî 0 æîíè áîðî0 äàòîãî 0 ä'iäà. Òðí, Ëï÷. ÌÀÐÒÎ0 ÂÀ ÁÀÁÀ, åòí.
Ìiôi÷íà iñòîòà, ïîâ'ÿçàíà ç ïðèõîäîì âåñíè. Êîëè çàã0 ð'iéå ìàðòà òà
òåïëî, òà 0 êàæó÷, ùî ìàðòîâà áàáà ç êîçàìè éäå. ÑÍIÃÎÂÀ ÁÀÁÀ.
Çëiïëåíà çi ñíiãó ïîäîáà ëþäèíè. Òà0 êó 0 ä'iòè ó âîø0 êîë'i 0 áàáó çë'i0 ïèëè,
øî éøëàì òà íàïóäèëàñ'à. Ëï÷. Ëêö, Òð. ÑËIÏÀ ÁÀÁÀ. Äèòÿ÷à
0 áóäå
ìàëèí0 êîéå
ãðà, ïiæìóðêè. Ìæ. ÁÀÁÀ-ÊÎÐÎËÀÁÀ, æàðò. Äóæå ñòàðà, õóäà
áàáà. Iäå äî íàñ 0 áàáà-êîðî0 áà. Ãðø. (There is the rst of 10 dictionary
entries for BABA in [Grytsak]).
2. Á
àáêà, -è, -îþ(-îé). 1. Âîðîæáèòêà, çíàõàðêà, øåïòóõà, ÷àðiâíèöÿ. [òà| êè
ñòಠõà| ðîøè / éàê ñ'iì | áàáîê ïîøåè ï| òàëî] Ïðîñ. Äèâ. ùå â
¹äüìà. 2.
Ðiæêè, ãðèáêîâèé ïàðàçèò ó êîëîñêàõ. [| æèòî ïî| éiëà | áàá∧ êà] Áiëëóö,
Ïiäã. Äèâ. ùå ÷îðí
óõà. [SUSG].
|
|
3. ÁÀÁÀ: áàáà ðî çîäðàíà (34) `íåöíîòëèâà ìîëîäà'. ÁÀÁÈ | áàáè (60), áà| áè
(26, 52, 60, 62, 70, 76, 94, 101) 1. `ïðèñóòíi íà âåñiëëi æiíêè' (101);
2. `ñâàøêè ìîëîäîãî / ìîëîäî¨' (70, 94); 3. `æiíêè, ùî âèãîòîâëÿþòü
âåñiëüíå ïå÷èâî' (62, 94); 4. `æiíêè, ÿêi âèãîòîâëÿþòü âåñiëüíi ñòðàâè'
(26, 76); 5. `æiíêè, ùî âèêîíóþòü âåñiëüíi ïiñíi' (52, 60, 94); 6. `æiíêè âiä
ìîëîäî¨, ùî ïðîäàþòü ¨¨ ïðèäàíå ðîäè÷àì ìîëîäîãî' (52); 7. `æiíêè,
ùî çàïëiòàþòü êîñó ìîëîäié íàïåðåäîäíi âåñiëëÿ' (60). ²| äàë i áà| áè (94)
`æiíêè, ÿêi âèãîòîâëÿþòü âåñiëüíå ïå÷èâî'. [Ìagrytska].
4. ÁÀÁÀ, íàç. ìí. ∼áè [ Ì-öü Ìàò. III, 33], ∼á
è [ Ê-â, Î-â], ðîä. ìí. ∼á²
[ Á-òå, Á-ëÿ, Æ-íÿ, I-¨, Ê-â, Ë-íà Â., Ë-öü, Ï-ï, Ï, Ø], ∼á [ Æ-íÿ] 1.
`ñòàðà æiíêà'; 2. `ìàòè áàòüêà àáî ìàòåði'; 3. ïåðåí. `ñòàðà êîðîâà' [ Ïàñ.
129] ; 4. äèâ. áàá
èöü [ Î-â], ïîð. ïîë. baba; 5. `ñêëàäîâà ÷àñòèíà êðîñåí'
[ Æ-í Êîá. 41, ß-öÿ]. - Á
àáà ïîä'áíà íà ïð
àíèê i âîí
à âõ
îäèò ó ä'äà: éå
òàê
èé êîë
îê, ø÷î çàïèõ
àéåñ'i çâåðõà ² á
àáó, ø÷îá
è øïàíóâ
àëî ïð'àæó
[ ß-öÿ] ; çàã. [ Æ-í]. Ñòîé
ò á
àáà ó êóò' ó çåëåí'iì êàáàò', õòî éié¨
ïîð
óøèò, ãóë'
àòè ç íå² ì
óñèò (ìiòëà) [Înyshkevych].
Every source has its own structural elements, located according to their own sequence,
designed with an individual set of fonts and their descriptions. Experience shows that it
is problematic to process two dictionaries by different authors using the same linguistic
algorithm, oriented towards extracting the maximum detail for a vocabulary entry. It
is therefore necessary to develop a separate algorithm for the processing of dictionary
Electronic Lexical Card Index for the Ukrainian Dialects
139
entries for every new dictionary, or to make a stage-by-stage restructuring of linguistic
information.
Therefore, three versions of the structure of electronic cards have arisen dependent
upon the degree of complexity: simple, standard and extended. A standard version
involves a restructuring of the dictionary entry according to the following levels: 1.
REGISTER UNIT (WORD); 2. Interpretation 1; 2.1. Forms of word (through comma);
2.2. Grammatical descriptions; 2.3. Stylistic descriptions; 2.4. Examples; 2.5. Phraseology; 2.6. Localization; 2.7. Other. Preparation of the extended version of a dictionary
entry involves the detailed realization of the explanatory part of a dictionary entry,
addition of new dictionary material, and the marking of semantic (thesaural) relations
between units of the database. At this stage of the work the simple version of card
was selected as being the closest to the concept of card index and the fastest to fulfil
practically, since it consists of two elements, marked out identically by all authors:
register unit and teh semantic part of the dictionary entry. The subsequent processing
of the card index is envisaged in order to broaden its structure (forming the extended
electronic card) and to improve the system of search and sorting of linguistic data.
The other issue is the software representation of information in the database and
the design of its interface. We made the structure of ELCIUD consist of six connected
tables and a software interface (figure 1). The window of the card index contains a
search engine, alphabetical index, editorial menu, menu for import and print, marker
for number of register units, alphabet and index sorting form, window for a dictionary
entry and form of dictionary entries sorted by their authors.
Fig. 1. The window of ELCIUD
ELCIUD is intended as an open system, so a system for addition of new information
is necessary. Information addition can be achieved in two ways in ELCIUD:
140
Pavlo Grytsenko, Olena Siruk, and Viktor M. Sorokin
1. addition of single linguistic units using the “Add” button of the menu
2. addition of data arrays by upload of two prepared files to the database through the
menu “Import”.
These files must have a precise structure. The first file must contain details of the
sources, written in accordance with the specially developed rules for preparation of the
dictionary materials (complete and short names of a dictionary) and lists (localization
of dialectal units after an author and after “Atlas of Ukrainian”, grammatical, stylistic
markers, additional signs used for phonetic transcription etc). The second file consists
of a dictionary, in which entries are given in a minimum (simple) version according to
whether there is selected: 1. a register unit (word) or 2. the semantic part of a dictionary
entry. The preparation of material in the format of standard and extended versions
of vocabulary entry is also allowed. The difference between them is in the degree of
detail of the dictionary material and these are not obligatory at this stage of the project
development, unlike the first (simple) version.
The menu “Import” gives a possibility to add new dialectal units to the card index.
It is recommended to upload preliminary prepared files to the database in the test mode
at first (if there is an error in the structure of these files, the program will warn about
it) and import of the tested file giving the numbers of added units. It is possible also to
specify the necessary number of added headwords, giving the additional possibility of
verification of data in the prepared file (figure 2).
Fig. 2. The window for “Import” of the program ELCIUD after verification
However, in spite of the double system of verification for added files, there is still
a need for the possibility of later correction of text and structural errors. An edit menu
has been developed for this purpose, which enables working both at the level of register
units and at the level of card texts. It is possible to correct text errors and edit font styles,
and also unite dictionary entries which have become mistakenly disconnected.
Electronic Lexical Card Index for the Ukrainian Dialects
141
The practical implementation of the system revealed problems with the upload of
data. In particular, there were:
1. loss of accents and space characters
2. improper representation of the phonetic transcription signs
3. chopping of the dictionary entry
In order to avoid the first two problems it is necessary to save documents in .rtf format
as web pages in Microsoft Word 2003. The third problem is solved by file testing
described previously or by editing already added dictionary entries using the program
menu (searching for bugs through index sorting and correction of contents using buttons
Äîäàòè/Âèäàëèòè (“Add”/“Delete”).
The problem of the complex localization mapping of dialectal units also arose the list of localizations and the card index exist separately in the database. Connection
is complicated because one card unites several meanings and several briefly recorded
localizations, respectively. A solution to this problem will be looked at in subsequent
stages of work with the database, at the same time as the detailed realization of the
dictionary entry.
The problem of search and sorting is related to the problem of localization mapping. Word search, author and word index sorting work in ELCIUD today. Enlargement
of the system of search and sorting will be possible with the detailed realization of card
structure.
In respect to the fifth problem, the prepared product (card index) is presented as
a database in the Microsoft Access format and additional software (the program Electronic dialectal card index (ELCIUD), developed in Visual Studio in the Ñ# programming language). Electronic cards are presented in electronic form only, but they can be
copied, saved as a file and printed.
5
Prospects for the development of electronic card
indexes
In the near future we plan to implement the following functions: possibility for printing
several entries; author screening to prevent the addition of identical linguistic information; web-presentation of the card index, to produce the technical editing of all card
indexes and to fill them with new register units.
The next and the most important task is the detailed realization of the card index
for the general improvement of work with the material. This will involve extending
the structure of the article, which will then consist of 21 elements. In accordance with
such a structure, a new approach to the database and program is needed, allowing for
transformation of the article in a new (extended) format. This will start a new stage
in our work: a new project for the transformation of card indexes into the format of
a complex electronic lexicographic system. This will relate the dialectal vocabulary
array to the literary language array of the Modern Ukrainian Language Dictionary (an
invariant field in the database) and will in future enable the structuring of linguistic
data according to a concept criterion (fields of synonyms, antonyms, hyperonyms, and
hyponyms).
142
Pavlo Grytsenko, Olena Siruk, and Viktor M. Sorokin
Data structuring is also a method of perfecting the system for search and sorting, and
for other valuable work with ELCIUD. It is the basis for the creation of database queries
using one or several criteria, which is of great value for complex linguistic research. At
this stage it is envisaged that there will be editing of the whole of the card index array,
not only at the level of the content of entries, but also at the level of the connection and
disconnection between register units.
6
Applications of ELCIUD
The project is intended for specialists and all those who are interested in the dialectal
layer of the Ukrainian vocabulary. An electronic card index can be used as a basis for
dialectological investigation, as an information system or for the purposes of education.
In combination with other linguistic software systems, ELCIUD can serve as a basis
for the analysis of the author’s style and for complex research into Ukrainian at the
different stages of its development.
7
Results
Within the framework of the project, the following tasks were performed:
1. theoretical principles of data preparation and the design of ELCIUD were drawn
up; the key concepts are the system, structure and formalization at all stages of the
work;
2. a database ELCIUD of six connected tables was produced in the Microsoft Access
format;
3. an interface and software was developed in Visual Studio with the use of the Ñ#
programming language;
4. a menu for reading and editing the database was constructed;
5. the minimum sorting of dictionary information is implemented (alphabetical, index
of headword and author sorting system);
6. we developed the possibility for enlargement of the database ELCIUD in two ways:
addition of register units one by one through the menu and addition of files which
contain the preliminary structured vocabulary entries;
7. a protocol was developed concerning ELCIUD data preparation by “virtual employees”;
8. the test version of the web interface for ELCIUD was developed. An online version
will be placed on the server of the Ukrainian Academy of Sciences or on the
linguistic portal MOVA.info;
9. future transformation of ELCIUD to into a structured electronic lexicographic product is planned.
Electronic Lexical Card Index for the Ukrainian Dialects
8
143
Sources
Arkushyn – Àðêóøèí, Ã. Ë. Ñëîâíèê çàõiäíîïîëiñüêèõ ãîâiðîê. Ó 2-õ ò. Ëóöüê
2000.
Grytsak – Ãðèöàê, Ì. À. Ñëîâíèê óêðà¨íñüêèõ çàêàðïàòñüêèõ ãîâiðîê. Ìàøèíîïèñ. Çáåðiãà¹òüñÿ â Iíñòèòóòi óêðà¨íñüêî¨ ìîâè ÍÀÍ Óêðà¨íè.
Magrytska – Ìàãðèöüêà, I. Ñëîâíèê âåñiëüíî¨ ëåêñèêè óêðà¨íñüêèõ ñõiäíîñëîáîæàíñüêèõ ãîâiðîê (Ëóãàíñüêà îáëàñòü). Ëóãàíñüê 2003.
Onyshkevych – Îíèøêåâè÷, Ì. É. Ñëîâíèê áîéêiâñüêèõ ãîâiðîê. Ó 2-õ ò. Êè¨â
1984.
SUSG – Ãëóõîâöåâà, Ê., ˹ñêîâà, Â., Íiêîëà¹íêî, I., Òåðíàâñüêà, Ò., Óæ÷åíêî, Â. Ñëîâíèê óêðà¨íñüêèõ ñõiäíîñëîáîæàíñüêèõ ãîâiðîê. Ëóãàíñüê
2002.
References
[1] Ãðèöåíêî, Ï. Óêðà¨íñüêà äiàëåêòíà ëåêñèêà: ðåàëüíiñòü i îïèñ (äåÿêi
àñïåêòè), [in:] Äiàëåêòíà ëåêñèêà: ëåêñèêîëîãi÷íèé, ëåêñèêîãðàôi÷íèé
òà ëiíãâîãåîãðàôi÷íèé àñïåêòè, Ãëóõiâ 2005, ñ. 523.
[2] Æèãî, Ï. Îïûò èñïîëüçîâàíèÿ ÝÂÌ â îáðàáîòêå ìàòåðèàëîâ ÎËÀ,
[in:] Îáùåñëàâÿíñêèé ëèíãâèñòè÷åñêèé àòëàñ. Ìàòåðèàëû è èññëåäîâàíèÿ. 1991-1993. Ñáîðíèê íàó÷íûõ òðóäîâ, Ìîñêâà 1996, Âûï. 21,
ñ. 5763.
[3] Êðþ÷êîâà, Î. Þ. Ýëåêòðîííûé êîðïóñ ðóññêîé äèàëåêòíîé ðå÷è è
ïðèíöèïû åãî ðàçìåòêè, [in:] Èçâåñòèÿ Ñàðàòîâñêîãî óíèâåðñèòåòà.
Ñåðèÿ Ôèëîëîãèÿ. Æóðíàëèñòèêà, ò. 7, âûï. 1, 2007, ñ. 3034.
[4] Êðþ÷êîâà, Î. Þ.; Ãîëüäèí, Â. Å. Òåêñòîâûé äèàëåêòîëîãè÷åñêèé êîðïóñ êàê ìîäåëü òðàäèöèîííîé ñåëüñêîé êîììóíèêàöèè, [in:] Êîìïüþòåðíàÿ ëèíãâèñòèêà è èíòåëëåêòóàëüíûå òåõíîëîãèè: Ïî ìàòåðèàëàì åæåãîäíîé Ìåæäóíàðîäíîé êîíôåðåíöèè Äèàëîã, âûï. 7(14),
Ìîñêâà 2008, ñ. 268273.
[5] Ëåêñè÷åñêèé àòëàñ ðóññêèõ íàðîäíûõ ãîâîðîâ. Ìàòåðèàëû è èññëåäîâàíèÿ. Ñàíêò-Ïåòåðáóðã 2005.
[6] Ëåñíèêîâ, Ñ. Â. Îïûò èíôîðìàòèçàöèè äèàëåêòîëîãè÷åñêèõ èññëåäîâàíèé. Ìîñêâà 1998.
[7] Òåð-Àâàíåñîâà, À. Â.; Êðûëîâ, Ñ. À. Ëåêñèêî-ãðàììàòè÷åñêèå áàçû
äàííûõ êàê èíñòðóìåíò äèàëåêòîëîãè÷åñêîãî îïèñàíèÿ, [in:] Êîìïüþòåðíàÿ ëèíãâèñòèêà è èíòåëëåêòóàëüíûå òåõíîëîãèè: Òðóäû ìåæäóíàðîäíîé êîíôåðåíöèè Äèàëîã 2006, Ìîñêâà 2006, ñ. 493498.
[8] Falińska, B. Leksyka dotyczaca
˛ hodowli na mapach Ogólnosłowiańskiego j˛ezykowego atlasu. Cz. 1-2. Białystok 2001.
[9] Gwary polskie. Przewodnik multimedialny, pod red. H. Karaś, [in:]
http://www.gwarypolskie.uw.edu.pl
[10] Kaś,
˛ J. Metody statystyczne w badaniach dialektologicznych, [in:] Acta Universitatis Lodziensis. Folia Linguistica, vol. 12, 1994, pp. 117–123.
144
Pavlo Grytsenko, Olena Siruk, and Viktor M. Sorokin
[11] Kretzschmar, W. Art and Science in Computational Dialectology, [in:] Literary
and Linguistic Computing, vol. 21, n. 4, 2006, pp. 399–410.
[12] Kretzschmar, W. Quantitative Areal Analysis of Dialect Features, [in:] Language
Variation and Change, vol. 8, n. 1, 1996, pp. 13–39.
[13] Nerbonne, J.; Kretzschmar, W. Progress in Dialectometry: Toward Explanation,
[in:] Literary and Linguistic Computing, vol. 21, n. 4, 2006, pp. 387–397.
[14] Siruk, O. Non-Technical Computer Thesaurus versus Specialized Computer Thesaurus, [in:] Metalanguage and Encoding Scheme Design for Digital Lexicography. Innovative Solutions for Lexical Entry Design in Slavic Lexicography.
MONDILEX Third Open Workshop, Bratislava, L’. Štúr Institute of Linguistics,
2009, pp. 177–182.
[15] Zar˛ebina, M. Słownictwo mieszkańców wsi Polski południowo-wschodniej (analiza statystyczna), [in:] Z polszczyzny historycznej i współczesnej, Rzeszów 1997,
pp. 35–50.
[16] http://www.upei.ca/methodsxii/html/e_speakers_papers.html (link
created on February 18, 2009 to the materials of the seminars conducted during
the work of the XII International conference Methods XII, held in August 2005 at
the University of Moncton (Université de Moncton)).
Inflectional Entropy in Slovak
Adriana Hanulíková1 and Doug. J. Davidson2
1
2
Max Planck Institute for Psycholinguistics, Nijmegen, Netherlands
Max Planck Institute for Human Cognitive and Brain Sciences, Leipzig, Germany
Abstract. Statistical measures of word frequency are used in psycholinguistic
research to characterize the psychological organization of the mental lexicon,
and the processes of retrieving, understanding, and learning words. More recently,
researchers have calculated statistics from corpora to gain insights into processing
of morphology, based on previous work on Serbian by A. Kostić and colleagues.
One such statistical measure - the inflectional entropy - has been shown to explain
processing costs in word recognition experiments. The inflectional entropy of a
word form is the amount of information carried by that inflected form, relative to
the statistical distribution of its inflectional paradigm. In this work, we investigate
whether it is possible to calculate measures like inflectional entropy for Slovak
using the Slovak National Corpus (SNK). This would allow us to compare Slovak
with other Slavic languages such as Serbian. The results will be useful for a
wide variety of psycholinguistic investigations of comprehension or production
of Slovak.
1 Introduction
Many psycholinguistic investigations have shown that the probability of a word has
a strong influence on measures of performance (for a recent review see Balota, Yap,
& Cortese, 2006). This is true for a wide variety of tasks, such as word recognition,
judgement tasks, or picture and word naming. For example, one of the most commonlyused tasks is the lexical decision task. In this task, the time it takes to judge whether a
singly-presented word occurs in a language is measured. Response times in this task are
faster for more common words relative to less common words (Whaley, 1978). Since a
Slovak word like ‘škola’ (book) is used more often than a word like ‘pštros’ (ostrich),
lexical decision times should be shorter for ‘škola’.
For the purposes of psycholinguistic studies, the probability (P r) of a word (w) is
often approximated, as in Equation 1, by estimating its unigram frequency count F (w)
in a sample of text or speech of size N (Baayen, 2001). These counts are typically derived from non-annotated corpora, which do not provide information about grammatical
classes or functions of the individual words.
Prw = F (w )/N
(1)
However, more recently researchers have incorporated variables related to morphosyntactic variation in the frequency estimates of words, based on annotated corpora (for review see Milin, Kuperman, Kostić, & Baayen, in press). This is especially
important for Slavic languages, which have richer inflectional morphology than the
146
Adriana Hanulíková and Doug J. Davidson
more-commonly studied West Germanic languages, and thus require more complex
probability models. In particular, work on Serbian by A. Kostić and colleagues has been
instrumental in demonstrating the influence of the inflectional form of a word on lexical
decision performance. Since this framework is the point of departure for the present
paper on Slovak, we will review some of their findings and conceptual distinctions
here.
Kostić (1991, 1995) found that the relative frequency of an inflected form within a
paradigm, as well as the number of grammatical functions or meanings of a word, was
correlated positively with lexical decision times for Serbian nouns. Their measures were
based on information theory, quantifying the amount of information that an inflectional
suffix provides, relative to its paradigm. More recently, Moscoso del Prado Martín,
Kostić, and Baayen (2004) found that lexical decision times for Dutch nouns were positively correlated with inflectional entropy. Inflectional entropy increases in a paradigm
when there are more inflectional variants possible, and/or when the variants have similar
probabilities. The key observation of this previous work is that the statistical distribution of word forms within an inflectional paradigm can be factored into two parts: The
contribution provided by the stem, and the contribution conveyed by the exponent (i.e.,
suffix). This is illustrated below in Table 1, which shows a probability model for the
Slovak feminine noun ‘škola’ (school), constructed in a similar way to Milin et al.
(2009, in press). The columns provide information on the surface frequencies F (we )
(per million) and surface relative proportions P rπ (we ) = F (we )/F (w), where F (w)
is the sum of all F (we ).
we
F (we ) P rπ (we )
škol-∅
211
0.09
škol-a
197
0.08
škol-u
248
0.10
škol-i,y
976
0.39
škol-e
598
0.24
škol-ou
66
0.03
škol-ám
15
0.01
škol-ách
146
0.06
škol-ami
22
0.01
Iwe
3.55
3.65
3.32
1.34
2.05
5.23
7.36
4.09
6.81
F (e) P rπ (e)
99396
0.11
139469
0.15
135748
0.14
312564
0.33
146867
0.16
68712
0.07
4890
0.01
17630
0.02
17576
0.02
Ie
3.25
2.76
2.80
1.59
2.68
3.78
7.59
5.74
5.75
Table 1. Probability distribution for the inflected noun škola.
The amount of information conveyed by the inflected words (we ) and exponents
(e) are calculated by applying the base -log2 transformation on the respective relative
frequencies of the different exponents, and the relative frequencies of the inflected
forms.
For example, the amount of information conveyed by the exponent ‘u’ (2.80) is
calculated from the probability of the exponent P rπ (e)
Ie = − log2 Prπ (e)
(2)
Slovak Entropy
147
where e = u (0.1439), estimated from the frequency of the exponent F (e) (135748)
relative to the sum of the frequencies of the exponents in the paradigm (942852)
Prπ (e) =
F (e)
Σe F (e)
(3)
There are also other statistical measures which represent properties of the entire
paradigm. The entropy of an inflectional paradigm, H, is calculated as
H = −Σe Prπ (we ) log2 Prπ (we )
(4)
For the values shown in Table 1 for ‘škola’, this is calculated as: H(‘škola’)= −[0.0851×
log2 0.0851 . . . 0.0089 × log2 0.0089], which amounts to 2.46. Informally, this index
captures the degree to which the paradigm is unevenly distributed over the different
forms.
In sum, these metrics characterize the contribution of stems and exponents to the
probability that a word form will occur. These measures are made practically possible
with the availability of relatively large morphosyntactically-annotated corpora such as
the Slovak National Corpus (SNK).
Here we want to investigate whether it is possible to calculate inflectional entropy
using the SNK, and if so, characterize how the results differ from previously reported
results from Serbian. These comparisons would support future empirical research on
word processing in Slovak, and help characterize differences between these two closely
related languages.
Number Case
Singular Nominative
Genitive
Dative
Accusative
Instrumental
Locative
Plural
Nominative
Genitive
Dative
Accusative
Instrumental
Locative
Serbian
planin-a
planin-e
planin-i
planin-u
planin-om
planin-i
planin-e
planin-a
planin-ama
planin-e
planin-ama
planin-ama
Slovak
planin-a
planin-y
planin-e
planin-u
planin-ou
planin-e
planin-y
planín-∅
planin-ám
planin-y
planin-ami
planin-ách
Table 2. Slovak and Serbian regular feminine inflectional exponents, illustrated with the noun
‘planina’ (meaning mountain in Serbian and plain in Slovak).
Despite the differences between surface exponents used in Serbian and Slovak (see
Table 2 above for an example), there are many similarities between the morphosyntactic
systems of Slovak and Serbian. Both languages have relatively complex inflectional systems, in which nouns are marked for number (singular and plural) and grammatical case
148
Adriana Hanulíková and Doug J. Davidson
(nominative, genitive, accusative, dative, instrumental, locative; the vocative is archaic
in Slovak and its status is disputed in Serb). In addition, the inflectional endings depend
on the gender of the noun (feminine, masculine, neuter) and the inflectional class.
Given such similarities, we would expect that statistical distribution of the Serbian
and Slovak terms would be similar. If we take the example of a base-level term used in
Milin et al., such as ‘žena’ (woman), we should observe a similar statistical distribution
as their Slovak counterpart ‘žena’, because they would be expected to have a similar
distribution of grammatical functions and meanings. If this is the case for most of the
terms in Slovak, then many of the psycholinguistic results obtained from the study of
Serbian should also generalize to Slovak.
On the other hand, there might be some reasons to expect differences between these
(and also other Slavic) languages. First, some of the basic-level terms in the two languages have different meanings, gender or inflectional class. For example, the primary
meaning of ‘planina’ (mountain in Serbian) does not correspond to the same meaning
as its Slovak counterpart ‘planina’ (plain in Slovak). Second, the statistical estimates
for Serbian are based on a sample of text, as is the case with all statistical parameter
estimates. It may be the case that the parameter estimates for a given measure like
inflectional entropy will be conditioned on the data source. This would suggest that the
Slovak and Serbian parameter estimates could be different, either due to real differences
in the usage of the two languages, or to differences in the samples used to estimate the
parameters.
We hypothesized that the factors governing the paradigm distribution of nouns in
Slovak and Serbian would be similar. We predicted that the measures of inflectional
entropy and paradigm entropy of Slovak and Serb would therefore also be similar.
2 Method
For a global comparison with Serbian results, we created two figures as in Milin et al.
(2009:55). We made a query from the SNK for all feminine and masculine nouns in all
respective cases and numbers. We then extracted statistical information for all feminine
exponents. Masculine nouns were not further analyzed. Milin and colleagues focused
on dominant regular inflectional subclasses in their paper; we consider all feminine
exponents. Note that hy, ii exponents were not computed separately, since in modern
Slovak they both express the same phoneme /i/. The function of hyi is to indicate that
the preceding sound is not palatalized.
For the comparison of inflectional entropy between the two languages, we selected
words from the word list provided in Milin et al. (2009) for which there was (almost)
complete form overlap with their Slovak counterparts, and used these for the query in
the manually morphologically annotated subcorpus r-mak-3.0 from the SNK. For the
analysis, we used only those words that were present two or more times in the SNK
sample, and we did not include diminutives. The frequencies and relative frequencies
of inflected variants and inflectional exponents were computed in the same way as in
Milin et al. (2009) and as described earlier in the Introduction.
Slovak Entropy
149
Fig. 1. The relative frequencies of feminine and masculine nouns for Slovak according to case
and number.
Fig. 2. Relative frequency of feminine nouns in Slovak according to inflectional suffix.
3 Results and discussion
Figure 1 shows for each case-number combination the distribution of relative frequencies within each inflectional class (here, the masculine and feminine nouns). Except for
the values of relative frequencies, the picture is almost identical to the Serbian results.
This is a good example of how different corpora can still be representative with respect
to morphological aspects of language use, irrespective of whether it is of a smaller
or larger size. Figure 2 plots the relative frequency of individual exponents within the
feminine inflectional classes. These are also considerably similar to Serbian.
Now we turn to the question, whether the inflectional entropy of individual cases is
comparable as well. Table 3 shows the inflectional entropy, H, calculated for the words
we selected from the Serbian lists. The average entropy for Slovak (µ = 1.70), in this
sample, was less than Serbian (µ = 2.11), t(18) = 2.011, p = 0.059. The correlation
between the two samples was relatively low, r = 0.2. This result would suggest that the
deviation from the paradigm pattern is, on average, greater for Serbian than for Slovak.
150
Adriana Hanulíková and Doug J. Davidson
Slovak
kniha
rieka
búrka
tráva
brigáda
fabrika
škola
náuka
ruža
stanica
ulica
dolina
duša
ryba
sila
potreba
vŕba
hlava
hviezda
H
2.63
2.28
1.30
1.52
0.65
0.86
2.46
0.88
1.24
1.72
3.04
0.59
2.36
1.71
3.27
2.74
0.24
0.80
2.01
Serbian
knjiga
reka
bura
trava
brigada
fabrika
škola
nauka
ruža
stanica
ulica
dolina
duša
riba
sila
potreba
vrba
glava
zvezda
H
2.17
2.22
2.23
2.23
1.89
2.12
2.20
1.98
1.90
2.05
2.39
2.43
2.28
1.79
2.03
2.13
1.86
2.34
1.83
Table 3. Comparison of Slovak and Serbian word pairs.
This result suggests that despite the similarities between Serbian and Slovak, their
inflectional entropy differs. However, several caveats should be kept in mind. This
comparison was based on a relatively limited number of words, and in order to maintain
strict comparability, we only examined words with overlapping surface forms. Despite
this overlap, preferences for certain terms, or differences in meaning in the respective
languages, could lead to differences in the frequencies of some terms. Future work
could examine larger samples, and other inflectional classes.
Despite the small sample, the results offer some suggestion that individual measures
of entropy are needed for each language, even for languages as typologically similar
as Serbian and Slovak. In practical terms, it appears that the use of morphologicallyannotated corpora are very helpful for calculating these measures for each language. A
useful framework for future comparisons of Slavic languages (or other languages that
have similar inflectional classes) might include measures like inflectional entropy in
order to guage the similarties and differences between languages.
4 Summary
In this paper we have described how inflectional entropy can be estimated from the
Slovak National Corpus. The obtained estimates were compared to results reported
previously for Serbian. The results showed that overall, the distribution of feminine and
masculine inflected nouns (grouped according to case and number) is almost identical
for both languages. The comparison of relative frequencies for feminine nouns, grouped
Slovak Entropy
151
by inflectional suffixes, showed a considerable amount of similarity with Serbian, despite the differences in suffix forms. Given this outcome, we expected inflectional entropy measures for a selected number of Slovak and Serbian (high frequency) nouns
to be comparable. However, the results showed that the estimates differ. This implies
that morphologically-annotated corpora could be very useful for cross-linguistic comparisons.
5 Acknowledgements
We would like to thank the producers of the Slovak National Corpus for making their
work widely available. This work was made possible by the support of the Max Planck
Gesellschaft. Both authors contributed equally.
References
Baayen, R. H. (2001). Word frequency distributions. Dordrecht: Kluwer.
Balota, D. A., Yap, M. J., & Cortese, M. J. (2006). Visual word recognition: The journey from features to meaning (A travel update). In M. Traxler and M. Gernsbacher
(Eds.) Handbook of Psycholinguistics, 2nd Edition. Pp. 285–375. Amsterdam: Elsevier.
Kostić, A. (1991). Informational approach to the processing of inflected morphology:
Standard data reconsidered. Psychological Resarch, 53, 62–70.
Kostić, A. (1995). Informational load constraints on processing inflected morphology.
In L. B. Feldman (Ed.) Morphological Aspects of Language Processing. Pp. 317–
344. New Jersey: Lawrence Erlbaum Inc. Publishers.
Milin, P., Kuperman, V., Kostić, A., & Baayen, R. H. (in press). Words and paradigms
bit by bit: An information-theoretic approach to the processing of inflection and
derivation. In J.P. Blevins & J. Blevins (Eds.), Analogy in grammar: form and
acquisition. Oxford University Press: Oxford.
Milin, P., Durdevic, D. F., & Moscoso del Prado Martín, F. (2009). The simultaneous
effects of inflectional paradigms and classes on recognition: Evidence from Serbian. Journal of Memory and Language, 60, 50–64.
Moscosos del Prado Martín, F., Kostić, A., Baayen, R. H. (2004). Putting the bits
together: An informational theoretical perspective on morphological processing.
Cognition, 94, 1–18.
Slovenský národný korpus, – r-mak-3.0. Bratislava: Jazykovedný ústav. L’. Štúra SAV
2008. http://korpus.juls.savba.sk.
Whaley, C. P. (1978). Word-nonword classification time. Journal of Verbal Learning
and Verbal Behavior, 17, 143–154.
Exploring Derivational Relations in Czech with the
Deriv Tool⋆
Dana Hlaváčková1, Klára Osolsobě2 , Karel Pala1 , and Pavel Šmerk1
1
Faculty of Informatics, Masaryk University
Botanická 68a, 60200 Brno, Czech Republic
{hlavackova,pala,smerk}@fi.muni.cz
2 Faculty of Arts, Masaryk University
Arna Nováka 1, 602 00 Brno, Czech Republic
[email protected]
Abstract. The aim of this paper is to present a tool for testing automatic word
derivation for Czech. The derivation of some word formation types in Czech is
in high degree regular. It can be described by formal rules. A new version of the
web interface Deriv working with the morphological analyzer ajka enables us
to formulate more complex word formation rules and to test more complicated
cases of derivational relations. The second issue touched extensively in the paper
are the types of derivational relations and their semantic classification. We have
proposed 14 semantic classes for suffixes and 11 for prefixes. The tool Deriv
helps considerably in establishing semantics of the derivational relations.
1 Introduction
In the paper we present the recent results obtained in the area of automatic processing of
Czech morphology (inflectional and derivational) in cooperation between the Department of Czech Language, Faculty of Arts, Masaryk University (K. Osolsobě) and the
Centre for Natural Language Processing, Faculty of Informatics of the same university
(D. Hlaváčková, K. Pala, P. Šmerk).
In section 2 we briefly characterize existing tools for automatic processing of the
formal morphology of Czech (inflection) as well as the derivational morphology (word
formation). In the following section we explain how changes on the formal level correspond to semantic relations between related words (motivation relations). In section
4 we show how the relations between word base and derived word can be described
formally. We also offer some examples showing how the particular rules can be formulated and used in the new version of the software tool Deriv which allows us to test
these derivational rules on a large machine dictionary of Czech word forms. In the next
section we present some obtained results, especially how the new features in the tool
Deriv allows to formulate more general rules without a significant loss of precision.
Finally, in section 6 we summarize the results obtained so far in the course of testing
the software tool Deriv and outline the further possibilities of its use.
⋆
This work has been partly supported by the Ministry of Education, Youth and Sports within
the National Research Programme II project 2C06009 and project LC536.
Exploring Derivational Relations in Czech
153
2 Tools for (Czech) morphology
We work with two software tools: the first one is earlier developed morphological
analyzer ajka [13] and the second is newer web derivational interface Deriv [14] which
uses ajka and its data (however the tool itself is language independent).
2.1 Morphological analyzer ajka
The first of our tools that is able to process Czech word forms (both recognition and
generating) as well as some regular derivations is the morphological analyzer ajka3 .
It was developed in the Centre for Natural Language Processing at the Faculty of
Informatics, Masaryk University. The tool works with the formal description of the
Czech inflectional paradigms that was developed by Osolsobě [8] and with the list of
Czech stems (approx. 400 000 items).
2.2 Interface Deriv
The second tool, which we are going to present in detail, is a web interface Deriv4 .
It is an interactive tool able to process derivational relations in Czech. According to
the defined rules it can generate lists of words from the Czech word-form list (more
than 60 000 000 items). It is based on the assumption that it is possible to extract
n-tuples of the items from the wordlist of ajka (or from corpus) that meet two following
requirements:
– all members can be found in the wordlist;
– each member of the n-tuple meets the underlying hypothesis.
The list of n-tuples generated automatically in this way is subsequently manually checked
and the list of correctly derived n-tuples and the list of the exceptions (over-generated
cases) is obtained. The Deriv tool works in three steps:
– by using a grammatical tag and regular expression both appropriate strings and
substitutive rule are defined;
– the respective pairs base form/derived form are automatically generated;
– the obtained list of the n-tuples is further processed manually.
Incorrectly derived word-form are also collected as negative examples for the next
iteration of the more complex rules.
3 The types of the derivation relations and their semantics
The semantic relations (traditionally characterized as motivation) can be captured by
formal rules based on regularities on the level of graphical realisation of phonemes,
combining affixes and parts of speech. They exploit Czech traditional word formation
grammar [3], and semantic classifications worked out for Czech WordNet and lexical
database VerbaLex [5].
3
4
More information at http://nlp.fi.muni.cz/projekty/ajka, the interactive version is
accessible at http://nlp.fi.muni.cz/projekty/wwwajka.
More information at http://deb.fi.muni.cz/deriv.
154
Dana Hlaváčková et al.
3.1 Semantics of the derivational relations
In our view, there is a need for a more detailed and better semantic specification of
the derivational relations (D-relations) in comparison with the traditional labels that we
consider rather broad and too general.
On the ground of the previous research [5] we offer a collection of the 15 derivational relations (see also [2] and [6]). We try to capture the main relations and leave
aside either the marginal ones or the ones that are not using strictly suffixation and
prefixation, i.e. compounds.
They can be tentatively grouped in the following way:
– role D-relations in which a relation between two POS expresses a semantic role,
e.g. Agentive, Patient, Instrument, Location, etc.;
– gender, diminutive and augmentative D-relations;
– D-relations denoting various kinds of properties, some of them can be considered
deverbative (first member of the pair is always a verb) and the remaining ones exist
between nouns and adjectives (also possessives) and adjectives and adverbs;
– prefix D-relations consisting only of the pairs verb – prefixed verb, their meanings are partly related to the verb classes, e.g. verbs of motion, verbs of drinking
and eating, or verbs of weather, etc. This area represents a challenge for further
research.
From the first 14 D-relations using suffixation the basic and most productive ones
have already been integrated into Czech morphological analyzer ajka together with
their semantic labels. The 15th one is, in fact, a complex group of the 11 relations
existing between verbs and using prefixation only. They are mentioned briefly below.
The indicated types of the D-relations can be further characterized in the following
way:
– D-relations expressing semantic roles similar to deep cases, they exist between verb
– noun pairs and use suffixation (numbers in the brackets denote frequency of the
pair in the corpus SYN20005, which is part of Czech National Corpus [1]):
• der-agentive: učit – učitel (8639 – 9924), teach – teacher (pair verb – noun);
• der-patient: trestat – trestanec (1636 – 301), punish – convict (pair verb –
noun);
• der-instrument: ukazovat – ukazovátko (11 831 – 61), point – pointer, fescue
(pair verb – noun);
• der-location: letět – letiště (3213 – 6636), fly – airport (pair verb – noun);
– D-relations denoting gender derivating of feminine nouns from their masculine
counterparts and and D-relations expressing diminutivity and augmentation, they
exist between noun – noun pairs. In Czech deminutives occur also in triples (suffixation):
• der-gender: student – studentka (15 608 – 1260), student – she-student (pair
noun – noun);
5
A detailed information on the structure of corpus SYN2000 can be found at
http://www.korpus.cz/english/struktura.php
Exploring Derivational Relations in Czech
155
• der-dimin: dům – domek – domeček (46 485 – 5118 – 606), house – small
house, very small house or house one likes (triple noun – noun – noun);
• der-augm: dub – dubisko (752 – 24), oak tree – huge and strong oak tree (pair
noun – noun);
– D-relations denoting action and property (suffixation) – their typical feature is that
both members of the relation refer to the same meaning and they differ only in
the part of speech, for instance, in the pair učit – učení (teach – teaching) the
action is denoted by a verb and deverbative noun, respectively. In the other pairs
both members denote property and the difference consists again only in the part of
speech:
• deriv-action: učit – učení (8639 – 2877), teach – teaching (pair verb – deverbative noun);
• deriv-property: učit – učený (8639 – 409) teach – learned (pair verb – deverbative adjective);
• deriv-property: učený – učenost (409 – 63) learned – learnedness (pair adjective – noun);
• deriv-property: učený – učeně (409 – 33) learned – learnedly (pair adjective –
adverb);
• deriv-property-possessive: učitel – učitelův (9924 – 56) teacher – teacher’s
(pair noun – adjective);
– D-relations exploiting prefixation represent a separate complex group in which we
observe only the pair verb – verb. They express meanings depending heavily on
the meanings of the verb stems they occur with. They denote a number of different
semantic relations such as various sorts of motion, time and location (see below),
intensity of action, inchoativity, iterativity, additivity, distributive action, obligation, result and possibly some others. Here we offer their preliminary classification
which obviously requires a further investigation. It is based on the list of 14 Czech
basic (primary) prefixes that can be usually found in Czech grammars [12]: do- (to),
na- (on, at), nad- (above, up), od- (from, away), pro- (for, because), při- (by, at),
pře- (over), roz- (over), s-/se- (with, by), u- (at, near), v-/ve- (in, up), vy- (out, off),
z-/ze- (of, off), za- (over, behind).
Verbal D-relations using prefixes can be further characterized:
– motionI: deriv-mot-to: motion to the point or place, e. g. jít – přijít, (go – come),
letět – přiletět (fly – arrive by plane);
– motionII: deriv-mot-to-iter: iterative, repeating motion to a point or place, e. g.
přicházet – přicházívat (be coming – be coming repeatedly);
– motionI: deriv-mot-from: motion from a point or place, e. g. jít – odejít (go/walk –
leave by going/walking);
– motionI: deriv-mot-from-iter: iterative motion from a point or place, e. g. odcházet
– odcházívat (leave by walking – leaving repeatedly);
– motionI: deriv-mot-over: motion across a point or place, e. g. brodit – přebrodit
(wade – wade through);
– motionII: deriv-mot-under: motion under a point or place, e. g. letět – podletět (fly
– fly under);
156
Dana Hlaváčková et al.
– timeI: deriv-compl-act-: to complete an action (with regard to any verb), e.g. letět
– doletět (fly – finish flying);
– timeII: deriv-t-act-iter: to complete an action iteratively, e.g. tancovat – dotancovat
– dotancovávat (finish dancing – finish dancing repeatedly);
– obligation: deriv-oblig: to perform an action as an obligation pracovat – odpracovat
(work – work off);
– additivity: deriv-addit: action expressing adding koupit – přikoupit (buy – buy more);
– distributivity: deriv-distrib: to perform an action in a distributed way ztratit – poztrácet (lose – lose little by little), i.e. lose successively particular objects one after
another;
– result: deriv-result: to perform an action with its result vařit – vyvařit (boil – boil
away);
– high intensity: deriv-high-intens: to perform an action more intensively vařit –
navařit (cook – cook a lot of sth);
– low intensity: deriv-low-intens: to perform an action with low intensity pracovat –
popracovat (work – work a little, for a while).
The list is preliminary and in our view it includes just the main and most typical
meanings expressed by the primary prefixes do- (to), od- (from), na- (on, to), po- (after),
pod- (under), pře- (over, across), při- (to), vy- (out, from). We give examples of all
meanings mentioned above but not necessarily with all prefixes. It can be seen that
meanings related to motion and time can be further subclassified but we certainly have
not captured all that can be related to the verbs of motion. In this respect the more
detailed investigation is necessary. We also consider only the verbs of motion with two
arguments, i. e. verbs with an Agent causing motion to a Location. Verbs with an Agent,
moved Object, e.g. nést knihu domů – přinést knihu domů (carry the book home – fetch
the book home) and Location (home) still have to wait for the more detailed analysis.
In the list above we have included iterativity relation as well because of its regularity in
Czech though iteratives are not derived with prefixes but with alternations in the stems
using infixes -áva-, -íva-, -ova-. On one hand, including the iterativity relation may
seem to complicate the description but on the other, its regularity allows us to handle it
almost automatically. It has to be remarked that the iterativity relation is semantically
close to the aspect relation, that is why some authors speak about third aspect though
iteratives are imperfective by definition.
We are well aware that the D-relations using prefixation can be organized differently
as it is typical for any kind of semantic classification. We are attempting to find one of
the possible solutions.
The need for describing semantics of the D-relations comes from the fact that Drelations are directly accessible for language users and make them able to understand
meaning relations existing in the the text. Their formal description is a necessary condition for implementing more intelligent semantic search in the applications like Semantic
Web.
4 Word derivation rules for Deriv
Word derivation, i.e. forming new words from the corresponding word bases (words),
can be formally described as an operation over strings of the characters (lemma, word
Exploring Derivational Relations in Czech
157
form, morphological tag). The derivational relations are typically regular and can be
described as formal rules for the Deriv interface. This allows us to reduce the machine
wordlist as well as to minimize the list of roots.
The very first versions of the tool Deriv only allowed formulation of rules based
on simple substitution of a final (or initial) substring of characters. Later the support
for regular expressions was added, so that the rules could describe also some simple
alternations in a more general manner.
4.1 New features of Deriv
To allow construction of even more general rules we have implemented the following
two features:
– the rules now can describe relations not only between lemmata but also between
arbitrary word forms (previous versions of Deriv were able to work with lemmata
only);
– we have added three general rules to describe alternations which cannot be directly
expressed by means of regular expression substitutions:
• palatalization of the final consonant;
• shortening of the last vocal;
• shortening of the last but one vocal.
Moreover, to allow further simplification of the rules we have added two another
features:
– utilisation of regular expressions6,of the programming language Perl 5.10, especialy of the construct \K which separates what is to be replaced and what have to
precede (in previous versions whole match was replaced);
– introduction of four shorthands $C, $V, $L, and $S for character classes “consonant”, “vowel”, “long vowel”, and “short vowel” respectively.
4.2 Structure of the rules for Deriv
The rules for Deriv have four parts:
–
–
–
–
regular expression which describes the tag of the first word of the relation;
regular expression which describes the tag of the second word of the relation;
regular expression which describes which part of the first word is to be replaced;
the replacing string.
The regular expressions use a syntax of Perl 5.10 programming language regular
expressions. The tags uses an atribut-value notation in which e.g. k1 stands for a noun,
k2 for an adjective, k5 for a verb, gM for a masculine animate, gI for a masculine inanimate, gF for a feminine, gN for a neutral, nS for a singular noun, c1 for a nominative,
c2 for a genitive, and mF for an infinitive7 . The application selects words with the tag
6
7
For full documentation see http://perldoc.perl.org/perlre.html
For full documentation see http://nlp.fi.muni.cz/projekty/ajka/tags.pdf (in
Czech).
158
Dana Hlaváčková et al.
matching the respective regular expression and then among these words it searches for
pairs in which the substitution of a string to be replaced in the first word of the pair for
the replacing string produces the second word of the pair. Only the pairs which meet all
conditions are returned to the user.
4.3 Some examples of the rules for Deriv
At first, an example of a simple rule – a particular class of adjectives can be derived
from the animate nouns with the suffix -í. The rule for automatic processing of pairs
like pták/ptačí (bird/of a bird) then is:
tag which selects the 1st word: k1gMnSc1
tag which selects the 2nd word: k2.*gMnSc1d1
string to be replaced: ák$
replacing string:
ačí
The rules are usually writen on one line in the form
ák$/k1gMnSc1 > ačí/k2.*gMnSc1d1
Examples of more complex rules A rule for an automatic processing of pairs like
řídit/řidič (drive/driver) can be formulated as follows (note the (.[aeěi]) and $1 constructs – whatever matching is between í and t in the first word, it is left in the second
word):
í(.[aeěi])t$/k5.*mF > i$1č/k1gMnSc1
Another interesting example is a rule for an automatic processing of pairs like Altaj/
altajský, where a change of capitalization of the first letter has to be described:
ˆ([[:upper:]].*)/k1gInSc1 > \l$1ský/k2.*gMnSc1d1
Examples of rules exploiting new features of Deriv The rule
$L$C\K[ay]$/k1g[FM]nSc2 > í/k2.*gMnSc1d1 /Mk
is a broad generalisation of the rule for pairs like pták/ptačí described above. /Mk at the
end of rule denotes an application of two of the general rules “palatalization of the final
consonant” and “shortening of the last vocal”, so that not only á and k will be shortened
and palatalized respectively, but also e.g. ů and h in bůh/boží (god/of a god) etc. The
first regular expression says that the first word of a pair has to end with a or y which
has to be preceded by a sequence of any long vowel and a consonant, but only the final
a or y will be replaced (because of \K in between). Moreover, note that the rule derives
from genitive forms (c2 in the first tag), not from nominative forms (lemmata).
Another example is a rule for an automatic processing of pairs like uchvátit/uchvatitel (usurp/usurper):
$L$C+$Vt\K$/k5.*mF > el/k1gMnSc1 /K
where /K at the end of the rule denotes an application of the general rule “shortening of
the last but one vocal”. Note the \K at the end of the first regular expression – it denotes
that only the words which ends with a long vowel, one or more consonants, any vowel
and t will be retrieved, but nothing will be replaced, so that only the addition of the
string el will occur.
Exploring Derivational Relations in Czech
159
5 Results
In the Table 1 we show how the new features of Deriv make possible to formulate
derivational rules more generally. The table compares sets of rules for four derivational
types which we studied in our previous work [4] (where one can found all particular
rules together with numbers of found pairs and over-generations for each rule and also
examples of the most of over-generations). The new rule sets have been constructed in
such a way that they find all correct pairs which are found by the original rule sets.
derivational type
example
agentive (-tel)
užívat/uživatel (use/user)
agentive (-č)
loupit/lupič (rob/robber)
purpose adjective (-í)
krýt/krycí (cover [v/adj])
generic possesives (-í)
husa/husí (goose [n/adj])
old version of Deriv
new version of Deriv
rules pairs
overgeneration
rules pairs
overgeneration
22
907
14 (1.5 %)
12
912
15 (1.6 %)
14
910
16 (1.8 %)
10
917
16 (1.7 %)
5
1746 3 (0.2 %)
1
1746 1 (0.1 %)
23
600
165 (27.5 %) 6
481
18 (3.7 %)8
Table 1. Comparison of old and new version of Deriv
In all cases there is a significant decrease of number of rules without any substantial
increase of over-generation. Also in all cases the new, more general rules found some
pairs which had been omitted before (mostly because of uniqueness of its alternations).
The Table 2 shows results of automatic processing of formally defined pairs verb
– agentive noun in Deriv and their manual classification9 . The figures for the first two
suffixes are the same as in the Table 1. But because the new features of Deriv are really
recent, we did not manage to work over the rule sets for all derivational relations for the
time being. Therefore to preserve mutual comparability of the figures in this table we
have to use older figures in all cases.
These results document an interesting fact that if the agentive nouns are derived
from a verbal stem (root + thematic infix, suffixes -tel and -č), the rate of over-generation
is significantly lower in comparison to cases where the agentive nouns are derived
directly from the root (suffixes -ce, -čí and -ec), even though these derivational types are
described by larger sets of rules. In other words, the derivation with the former suffixes
appears to be much more regular than the derivation with the latter ones.
8
9
This outstanding decrease of a rate of over-generation is caused by splitting feminines according to animateness: it is very essential information in the context of these rules as the generic
possesives can be derived only from the animate nouns.
More information on the derivation with suffixes -tel and -č can be found in [4, 9, 10], on the
suffixes -ce and -ec in [7] and on the suffix -čí in [11].
160
Dana Hlaváčková et al.
suffix
-tel
-č
-ce
-čí
-ec
example
užívat/uživatel (use/user)
loupit/lupič (rob/robber)
soudit/soudce (judge/judge)
mluvit/mluvčí (speak/speaker)
jezdit/jezdec (ride/rider)
rules
22
14
74
18
28
pairs
907
910
326
42
121
over-generation
14 (1.6 %)
16 (1.8 %)
51 (15.6 %)
11 (26.2 %)
45 (37.2 %)
Table 2. Results of processing of verb – agentive noun pairs
6 Conclusions
We have studied regularities of the Czech word formation and using the improved
version of the tool Deriv we have formulated and evaluated formal derivation rules for
different derivational relations. We have also paid attention to the semantics of the Drelations – the final task is to label the derivation rules also semantically and in this way
to make them usable in various NLP applications. Though a number of the individual
types have not been processed automatically yet, the presented results convincingly
show that Deriv allows to explore potential and limits of the formal description of the
derivational processes. The results can be used in NLP applications where semantic
search becomes the aim, but also e.g. for optimisation of existing software tools, especially the morphological analyser ajka and its stem list.
References
[1] (2000). Czech National Corpus – SYN2000. Institute of the Czech National
Corpus, Praha. Accessible at: http://www.korpus.cz.
[2] Azarova, I. V. (2008). Derivational semantic relations in RussNet. An oral presentation at the 4th Global Wordnet Conference. Szegéd, Hungary.
[3] Dokulil, M. (1962). Tvoření slov v češtině I (Word Derivation in Czech I). Nakladatelství ČSAV, Praha.
[4] Hlaváčková, D., Osolsobě, K., Pala, K., and Šmerk, P. (2009). Relations between
formal and derivational morphology in Czech. In Proceedings of Czech in Formal
Grammar 2009 Conference. Peter Lang Publishing Group. In print.
[5] Hlaváčková, D. and Pala, K. (2007). Derivational Relations in Czech WordNet.
In Proceedings of the Workshop BSNLP, pages 75–81, Prague. ACL.
[6] Koeva, S., Krsteva, C., and Vitas, D. (2008). Morpho-semantic Relations in
WordNet – a Case Study for Two Slavic Languages. In Proceedings of 4th Global
WordNet Conference, pages 239–253, Szegéd, Hungary.
[7] Kolářová, Z. (2009). Možnosti a meze automatické derivace – počítačové zpracování deverbativ na -ce a -ec (Potential and Limits of an Automatic Derivation
– Computer Processed Deverbatives with -ce and -ec Endings). Master’s thesis,
Faculty of Arts, Masaryk University, Brno.
[8] Osolsobě, K. (1996). Algoritmický popis české formální morfologie a strojový
slovník češtiny (Algorithmic Description of Czech Inflectional Morphology and
Czech Machine Stem List). PhD thesis, Faculty of Arts, Masaryk University, Brno.
Exploring Derivational Relations in Czech
161
[9] Osolsobě, K. (2008a). Formální pravidla derivace deverbativ na -č (Formal Rules
for Derivation of Deverbatives ending with -č). SPFFMU A, 56:121–135. Faculty
of Arts, Masaryk University, Brno.
[10] Osolsobě, K. (2008b). Propria (příjmení na -č) – problém automatické morfologické analýzy (Family Names with -č ending – a problem of automatical
morphological analysis). In Čornejová, M. and Kosek, P., editors, Jazyk a jeho
proměny (Language an it’s Metamorphosis). Prof. Janě Pleskalové k životnímu
jubileu, pages 205–216, Brno. Host.
[11] Osolsobě, K. (2009). Deriváty na -čí: gramatika, slovník a korpus (Derivates with
-čí: grammar, dictionary and corpus). SPFFMU A, 57. Faculty of Arts, Masaryk
University, Brno. In print.
[12] Petr, J. (1986). Mluvnice češtiny I (Grammar of Czech I). Academia, Praha.
[13] Sedláček, R. (2004). Morphematic analyser for Czech. PhD thesis, Faculty of
Informatics, Masaryk University, Brno.
[14] Šmerk, P. (2009). Deriv. Web application interface (in Czech), accessible at:
http://deb.fi.muni.cz/deriv.
On Epistemicity, Grammatical Person and Speaker Deixis
in Polish (Based on the Polish National Corpus)
Łukasz Jędrzejowski
Interdisciplinary Center “European Languages: Structures – Development - Comparison”
Free University of Berlin, Germany
Abstract. Examining data from the Polish National Corpus, I show to what
extent epistemicity is dependent upon grammatical person and speaker
deixis. The data corpus enables us to examine such instances of epistemic
coding, which are used infrequently. Thus, I pay attention to these
constructions, namely when the speaker and clausal subject collapse
referentially.
1 Introduction
The contribution investigates modal constructions in Polish, which express both
epistemicity and evidentiality. The discussion shows that if the speaker and clausal
subject collapse referentially, an epistemic meaning of a modal unit is excluded. In
pursuing this aim, I provide a description of the exotic modal musieć 'must' by
looking at its semantic properties with a particular focus on speaker deixis. After
characterizing the double displacement of musieć, I move on to discuss the
interpretations of the described modal expressions. Chapter 3 concludes the paper.
2 Modal verbs
If we take a look at languages that have a modal verb system, it may be said that
modal verbs function within the grammar as an exotic phenomenon. Having
considered Germanic languages, for instance, one may mention the preterite-present
forms, which in turn are foreign to Slavonic languages. Considering this fact, I have
adopted a universal definition of a modal verb proposed by Leiss (2009) (according
to Öhlschläger 1989):
Ein Verb ist dann als Modalverb to klassifizieren, wenn es neben der Grundmodalität
über eine zusätzliche epistemische Lesart verfügt. (Leiss 2009, page 6)
[A verb can be classified as a modal verb if it has at its disposal, in addition to
the root modality, another epistemic reading. Translation: Ł. J.]
Roughly speaking, Polish does not have at its disposal as many modal verbs as
German or Dutch. Nevertheless, we can find some Polish modal units which
correspond with our definition and which can be brought down to a common
denominator, i.e. to a polyfunctional one (móc 'can', mieć 'haben' and musieć 'must'
exemplify the most used modal units). In terms of polyfunctionality, I confine
myself to two main modal readings, namely deontic and epistemic. In what follows,
I scrutinize the modal verb musieć 'must' on the basis of the Polish National Corpus.
In order to go into detail, I will carefully examine constructions relating to the first
person singular, both masculine and feminine forms.
On Epistemicity, Grammatical Person and Speaker Deixis in Polish
163
2.1 The distribution of Polish musieć
It has been argued that the Polish modal verb musieć owes its existence to intensive
contact with German in the 12th century. Quite in line with Hansen (2000) we do not
know, however, whether Polish adopted it directly or via Czech. Nevertheless, it is
worth noticing that we can find in the 14 th and 15th century solely a deontic reading
of musieć, i.e. without its polysemy, and this deontic reading is still present in the
Contemporary Polish, as can be seen in [1]:
[1] Wszystko w życiu musiałem wypracować sam.
everything in life must [1.PS.SING.MASC.PAST.] work out
'I had to work out everything on my own in life.'
In the course of time musieć has obtained an additional meaning, namely an
epistemic one. What we usually have in mind when we talk about this phenomenon
is grammaticalization. The grammaticalization, as a kind of languages change, is a
process whereby lexical items become less lexical and thus they develop or take
over a grammatical function.
Epistemic modality evaluates proposition. It expresses an attitude of the speaker
towards the truth of a proposition, as is presented in [2]:
[2] Musiałem rabusiów spłoszyć podjeżdżając autem.
must [1.PS.SING.MASC.PAST.] robbers scare away driving up car
'I must have scared the robbers away driving up in the car.'
Musiałem both in [1] and [2] function as propositional operators, in contrast to
lexical verbs such as wypracować 'work out' or spłoszyć 'scare away'. They are,
however, some differences between these two modals. Musiałem in [1] belongs to
the circumstances; it governs not only the subject but also the bare infinitive,
whereas the modal verb in [2] does not. Musiałem in [2] involves the speaker's
conclusions about the circumstances. Moreover, one should not overlook the fact
that the modal in [2] has wide scope, while musiałem in [1] has narrow.
Now, let's take a look at the data from the Polish National Corpus. All
constructions consisting of the modal verb musieć in the past tense and the 1 st person
singular have been selected from the corpus and analyzed. Furthermore, the Polish
language, as the majority of Slavic languages, marks genders by means of person
ending. In terms ofthe research, I have found a total of 12341 tokens. Considering
the gender distinction, table 1 gives the outcomes of the masculine usage (i.e.
musiałem), - while table 2 shows the results of the feminine usage (i.e. musiałam):
deontic modals
epistemic modals
total
7748 (91,6%)
708 (8,4%)
8456
Table 1. Readings of musiałem in the Polish National Corpus
164
Łukasz Jędrzejowski
deontic modals
epistemic modals
total
3723 (95,9%)
162 (4,1%)
3885
Table 2. Readings of musiałam in the Polish National Corpus
The data shows that both interpretations occur simultaneously in the
Contemporary Polish. It must be pointed out, however, that if we take a look at the
statistical usage of musiałem and musiałam, the deontic reading takes precedence
over the epistemic. While over 90% of tokens are attributed to the deontic modality,
epistemicity seems to be overshadowed. In the next subsection, an attempt will be
made to characterize the epistemic constructions.
2.2 The double displacement
There is no doubt that such epistemic constructions, in which the speaker and
subject collapse, imply that the speaker refers both to himself and to his knowledge
status. He assesses a situation in which he was forced to act. At this point, it is interesting to note that we come across another grammatical category in addition to
epistemicity, namely evidentiality. A sentence like [2] points to a source of the
assessment and the marking of the source of the information of the statement is
known as evidentiality1. The coincidence of epistemicity and evidentiality is labaled
as the double deixis (cf. Abraham 2008b) or the double displacement (cf. Leiss
2009), and is reflected in Fig.1.:
proposition
subject
source of the information assessment of the information
truth assessor
speaker
propositional subject
illocutionary subject
Fig. 1. The double displacement of musiałem/musiałam2
1
2
It must be remarked that Polish does not have at its disposal pure grammatical markers of
evidentiality, as can be observed, for instance, in Bulgarian (cf. Sauerland and Schenner
(2007)).
The figure has been adopted from Leiss (2009).
On Epistemicity, Grammatical Person and Speaker Deixis in Polish
165
According to the Fig. 1., the clausal subject is divided into two parts: truth
assessor, who identifies source of the information, and speaker, who evaluates the
information. Taking into account this implication, it might be maintained that
propositional subject and illocutionary subject are conterminous, since they refer to
the 1st person. However, it is erroneous to follow this argument. In other words, the
illocutionary subject distances itself from the propositional subject, because it
indicates the lack of confidence in the truth of expressed proposition. This discursive
function of the I-perspective does not pertain to the 1st person. What it means, is that
the propositional subject has been shifted to the 3rd person. In this way, the speaker
acts as if he were another person, whom he relies upon3.
3 Conclusion
The weightiest result of this investigation is that as can be seen on the basis of the
Polish National Corpus, the full-fledged modal auxiliary musieć may be
characterized not only by deonticity but also by epistemicity, which in turn is bound
up with evidentiality. The implication of this findings is that musieć may code two
“grammatical” categories at the same time, on the one hand the source of information and its assessment, on the other. In both cases, we have to do with the 1 st
person which serves to assess proposition and therefore functions as a source of the
assessment. Additionally, one may say, that at first glance, the multiple deixis rests
upon the person in proposition, who might also be equivalent with the 1 st grammatical person. Having analyzed the Polish data, however, it can be concluded that
although the speaker refers to himself, he does not identify with the preposition
subject. The speaker distances himself from himself in such a way as though he
were another person.
References
[1] Abraham, W. (2008a). Aspektuelle und sprecher- bzw. personsgebundene
Bestimmungskomponenten deutscher Modalverben. In: Northern voices.
Essays on Old Germanic and related topics offered to Professor Tette Hofstra,
pages 250–269, Dekker, K., MacDonald, A., Niebaum, H. (eds.). Peeters.
[2] Abraham, W. (2008b). Illocutive force is speaker and information source
concern. What type of syntax does the representation of speaker deixis
require? Templates or derivational structure? Manuscript, University of
Munich.
[3] Bredel, U. (2002). “You can say you to yourself”. Establishing perspectives
with personal pronouns. In: Perspective and Perspectivation in Discurse, pages
167–180, Graumann, C., Kallmeyer, W. (eds.). Amsterdam/Philadelphia.
3
It is worth mentioning that the pronoun shift is also applicable to the pronoun you, as it has
been emphasized by Bredel (2000): “In empirical conversational situations, […], the usage
of the addressing pronoun du (you) as a linguistic sign does not always refer to the hearer,
but may also address the speaker” (Bredel 2002, page 167).
166
Łukasz Jędrzejowski
[4] Hansen, B. (2000). The German modal verb müssen and the Slavonic
Languages – The Reconstruction of a Success Story. In: Scando Slavica 46,
pages 77–93.
[5] Leiss, E. (2009). Drei Spielarten der Epistemizität, drei Spielarten der
Evidentialität und drei Spielarten des Wissens. In: Modalität. Epistemik und
Evidentialität bei Modalverb, Adverb, Modalpartikel und Modus, pages 3–24,
Abraham, W., Leiss, E. (eds.). Stauffenburg.
[6] Öhlschläger, G. (1989). Zur Syntax und Semantik von Modalverben des
Deutschen. Tübingen.
[7] Sauerland, U. & Schenner, M. (2007): Embedded Evidentials in Bulgarian. In:
Proceedings of Sinn und Bedeutung, 11, pages 495–509, Pui-Waldmüller, E.
(ed.). Pompeu Fabra.
A Russian EFL Learner Corpus from Scratch
Olga Kamshilova
Herzen State Pedagogical University of Russia
Saint-Petersburg, Russia
Abstract. Since Sylviane Granger’s research group findings (1998) Learner
Corpora (LC) have been looked upon as a valuable resource to pinpoint
widespread and recurrent learner errors, predominantly lexical. Online text tools
now available suggest working with grammatical issues, and their application,
handy enough, redirects a traditional linguist from applying to national or other
“ready-made” corpora to creating specialized “home-made” ones. Therefore
“from scratch” in this case is, by no means, ignoring the previous experience,
but developing a target-specific corpus structure. A written and spoken English
corpus from Russian students of intermediate level was set as a research
resource aimed at looking for other language events than those analyzed by
Sylviane Granger’s project, namely the range, complexity and frequency of
sentence patterns in school essays and dialogues. The desiderata at present are
comparative rather than didactic issues.
Keywords: special corpora, corpus format, EFL learner corpus, parsing in
corpora, basic structures, patterns, “structural” poverty
1 Introduction
LC, that have been widely in use since the end of 1990s, are large electronic
collections of written and, later, spoken texts produced by learners of English of
different levels such as International Corpus of Learner English (Louvain, Belgium),
Hong Kong Academic English, Michigan Corpus of Academic Spoken English,
English as Lingua Franca in Academic Settings (Tampere). They have provided
empirical means for two research traditions: that of second language acquisition
(SLA) theory (North America) and that of EFL teaching (Europe). The strong
tradition of LC analysis in Europe (S. Granger) revolutionized dictionary-making: the
use of corpus data transformed not only the process but also the product, and
dictionaries designed for learners of English have improved enormously (Granger,
Rundell 2007). One strong point in SLA theory is interlanguage (or IL, after Selinker
1972) – a “transfer” grammar of a learner language that displays systematicity and
deliberate choice rather than random error of other than native patterns, structures
which are supposed, with the growth of literacy, to be replaced by native-like ones. IL
analysis of higher-level EFL speakers, however, reveals that these preferences remain
(Hinkel 2003, Mauranen 2003). Moreover, the “transfer” English grammar of
speakers of different national languages tends to unification. This is actually the
grammar that ensures successful ELF (English as Lingua Franca) communication.
168
Olga Kamshilova
2 Hypothesis
IL analysis even at intermediate level may reveal the most frequent patterns and
structures that make up the core of EFL speakers’ actual grammar. These patterns and
structures are definitely native language oriented and only partly cover the list of
Basic English structures.
LC data analysis seems to be a most appropriate method to test the hypothesis. To
carry out the analysis there was a need for a special corpus that was launched as a
Master’s degree Program project (Камшилова, Колина, Николаева 2008). There
were several reasons for a new corpus. Firstly, the access to the existing LC data is
not free. Secondly, the available observations of learner text production are mainly
about vocabulary and pragmatic issues with little concern for morpho-syntactic
problems. Thirdly, most LC are compiled of advanced level texts, a level at which the
study modes and personal objectives as well as proper instruction provide a close to
native-like performance in written text production due to practiced patterns and
format. Fourthly, the existing LC do not include (with the exception of ICLE) English
texts produced by Russian students, so compiling a LC of Russian mother-tongue
learners seemed both opportune and challenging. Besides, the project aimed at
supporting the idea that with numerous free tools and techniques of analysis now
available a special corpus is becoming an efficient instrument in language research.
3 Corpus format
The target LC (SPb LC) compilation was based upon the Lerner Corpus design
criteria (Granger 1998:8) and use of free text tools now available that suggest working
with grammatical issues in accordance with the research objective.
The authentic EFL textual data were collected in Saint Petersburg schools in
November-December 2007. The contributors were 78 pupils of high school (forms 9–
11) with pre-tested language proficiency as intermediate (26%) and upperintermediate (74%)
The corpus size now is 38 122 words.
The language/text relevant criteria included medium, genre, topic, technicality and
task setting. The corpus contains two subcorpora of written texts (essays and personal
letters) and two subcorpora of oral monologues and dialogues (in scripts). The genres
and the topics were suggested by the school syllabus and the format of the most recent
innovation into the State Certification system of Russian School Education – ЕГЭ
(State General Exam). Task setting was deliberately different from the requirement of
S.Granger’s corpus: since the informants were quite familiar with the topics, the text
production was timed and a size limit for written text was set. No reference tools
(dictionaries or grammars) were used either.
The learner relevant information was collected in questionnaires and later shaped
the headers. The standardized header eventually included the following data:
A Russian EFL Learner Corpus from Scratch
169
Text Type: Essay
ID: Paul
Age: 17
Sex: female
School: 98
Form: 11
Mother-Tongue: Russian
L2: Swedish
Additional Language Training: no
Language level: Intermediate
Topic: Some people “live” in the Internet, they are real “websters”, while others never
navigate the Internet
Wordcount: 198 words
Date: 14.12.2007
The corpus was annotated with the help of free corpus managers (see the
Reference), the tools that produce such information as word counts, frequencies, and
collocation and syntactic patterns. Each text, as a result, is presented in the corpus in
three modes: original, parsed and error-tagged.
The choice of the corpus managers, parsers and error-taggers was guided by the
research objective (Разумова 2008). The authentic learner output data contain a high
rate of non-standard forms due to many spelling and grammar mistakes. That is why
some degree of caution should be exercised with standardized text retrieval software
(Granger 2002:15). The small size of the SPb LC is, in this way, rather an advantage,
since a manual correction can be applied.
The corpus tools include a corpus builder that is meant for the corpus completion.
4 Corpus findings: frequency list analysis
1
2
3
4
5
6
7
8
9
10
French
LC
the
of
to
a
and
is
in
that
it
be
Quebec
LC
the
to
I
a
of
and
in
that
is
it
SPb LC
all texts
I
to
and
you
the
a
is
it
in
like
SPb LC
Essay
to
the
and
I
is
a
it
you
people
have
SPb LC
Letter
I
to
you
and
in
the
my
a
very
is
SPb LC
Monologue
I
and
to
my
a
have
is
like
in
very
Table 1. Top 10 word forms in tree LC
SPb LC
Dialogue
I
you
to
the
like
and
we
go
it
think
170
Olga Kamshilova
A comparative frequency list analysis of three raw learner corpora – French learner
writing corpus (ibid:17), Quebec advanced LC (Cobb 2003) and the SPb LC – is only
a rough exploratory survey, but it provides some interesting perspectives.
It should be stressed again that the learner language proficiency varies in the
corpora from advanced in the French and Quebec LC to intermediate/upperintermediate in SPb LC. Besides, the task setting criteria were different: the
contributors to SPb corpus were set a time limit and did not use any reference
materials so that the text production was nearly spontaneous (except for the previous
class practice). Therefore, both the vocabulary and the sentence patterns presumably
reflect that actual language fund that the learners resort to in case of FL
communication. The rich vocabulary and developed sentence patterns trained in class
would give way to simple, common lexis and transparent structures. When compared
with what is vaguely known as Basic Grammar of English, these structures may,
hypothetically, shape the core of IL.
The top 10 in the compared corpora suggest that the task subjects in SPb corpus
were definitely 1st-person oriented, hence the first rank of “I”, but for the Essay
subcorpus. The reason is evident: the essay format, which results in like data for “to”
and “the” in the three corpora. The attraction is the high frequency rank of “and”, as
well as the verb forms “like”, “have” and “is” in SPb LC. As the concordance display
shows, the conjunction “and” is used to connect short and numerous clauses,
homogeneous parts, to start a sentence and to fill the pause in case of hesitation – all
that features in spontaneous speech production. The appearance of “have” and “like”
in The Top 10 List, which are with a minor exception used as notional ones (the Essay
subcorpus finds 109 hits for “have”, with only 4 auxiliaries and 14 modals) suggests
that SPb schoolchildren make wide use of the patterns I/WE <HAVE> Y and I/WE
<LIKE> C. The high rate of “think” in the Dialogue subcorpus is due to a standard
opinion phrase (I THINK) at the beginning (preferably) of the sentence. In the
comparable corpora the only verb forms that are found on the list are the forms of
“be” which can be either notional, link, modal or auxiliary. Though non-specified, this
rank suggests a greater variety of verb patterns and, consequently, sentence structures
in advanced learner output.
In SPb LC the verb form “is” was used 413 times. The cluster analysis
demonstrates 22 notional uses (there is, is in), 4 modal (is to), and 11 auxiliary ones
(3 for passive – is called/fried and 8 for progressive – is working/keeping/ making
etc.) So, the major use is in the link function in the pattern X IS Y.
5 Corpus findings: basic grammar patterns
Learner preference in pattern and structure choice was studied semi-automatically in
the parsed texts section of SPb LC with reference to concordance, frequency list and
collocation list. Following the hypothesis, the corpus data can pinpoint the most
frequent grammatical structures, used by foreign learners systematically and delibe-
A Russian EFL Learner Corpus from Scratch
171
rately, that shape the basic grammar of IL. The question is whether the list of
preferred structures bears any resemblance to the Basic English Grammar structures
and, if it does, what are the differences that bring to “structural poverty” and
“simplicity without elegance” (Hinkel 2003).
Though the idea of Basic Grammar is commonplace, there is more than one
attempt to its formal description (from the famous descriptive and structural models to
schemes for EFL teaching). The list taken as model in this paper is adopted from
Longman Grammar of Spoken and written English (1999):
•
•
•
•
•
•
S V A (Mary is in the house);
S V C (Mary is kind / a nurse);
S V Od (Somebody caught the ball);
S V Od A (I put the plate on the table);
S V Oi Od (She gives me presents);
S V (The child laughed).
A sample preliminary observation in the parsed corpus section provides
convincing evidence of the Basic structures employment in learner text production.
Each of the listed sentence models is found in learner texts. However, it is the
repetition of high frequency structures and the actual fill of the structure components
that imparts the non-native “accent”.
It is a commonly accepted fact that the models SVC and SVA (basically with be as
the link) prevail in non-native texts. The SPb LC data display the like tendency: the
relative frequency of predicates with be-link is 0.33.
V-components repertoire in S V Od, S V Od A, S V Oi Od, and SV patterns
demonstrates another feature of the actual learner grammar which is a deliberate
avoidance of perfective and progressive forms. This avoidance is felt by the native
speaker as a failure to nuance the aspect of the event/situation described, but for the
IL communication, it seems to be no obstacle. The preferable modal operator is can
(with absolute frequency in the corpus is 235 it ranks 20).
The frequency list analysis pinpointed the prevalence of have in essays and like in
monologues as compared to other notional verbs. It is worth considering the high rank
as the result of resorting to simple IL structures.
As the frequency list does not specify the part of speech, further analysis was
carried out with reference to the parsed texts section. The parser used for this purpose
was the demo version of ENGCG due to its being immune enough to the non-standard
learner text and for its convenient tags. The parsed texts analysis found 362 hits for
have in the whole SPb LC and 114 in the Essay subcorpus. Notional use of have was
found in 323 contexts with 114 contexts in the Essay subcorpus. The modal use is
high enough – 35 and 17 respectfully, while the use of auxiliary have (for Perfect
forms) is insignificant (4) and often inaccurate:
172
Olga Kamshilova
[I “have been falling in love” with “Counter-strike” for 6 years]1
The absolute prevalence of notional have suggests a wide use of the pattern S
<HAVE> Od. The suggestion is definitely backed by the cluster display that
highlights the usages of the type:
[I also have a cat and fish].
What seems quite special for the learner texts is the overuse of the pattern with the
Od position filled with nominalized forms, that is, they prefer a precast pattern to a
more conventional for native speakers nominal (be allergic, be free, be independent)
or verbal predicate (to communicate):
[I can understand when people don’t keep any animals because they have an
allergy]
[If you leave your childhood house you'll have your own life]
[If you have your own accomodation you also have a freedom]
[I enjoy to have communication with interesting people from different countryes].
Like is a frequent character in the whole corpus (359 hits), but since it is on the
Top 10 list in the Monologue subcorpus (139) the data of this part of the LC seem to
be of primary interest.
The ENGCG parser marks the notional use of like as @+FMAINV in 117
contexts. Combinations like I like studying are marked as complement structures –
@CS %CS (2). Twenty more word forms were marked as prepositions. The parser
failed only 10 times with the non-standard learner pattern (error) of the type [I very
like my family; I very like studying English; I very like to swim there].
The 117 contexts with the S <LIKE> C structure with Infinitive in the position of
C indicate that Russian learners do not differentiate the nuances of Infinitive and -ing
complements and systematically restore to the pattern with the Infinitive complement
similar to the Russian one (Я люблю читать). In some monologues the pattern is
numerous:
[I'm 15 years old. I was born in Krasnoyarsk and we moved to St. Petersburg
when I was 5 years. I live with my mother, father, sister and brother. My mother is a
doctor, father is an architect, my brother is 25 and he is a musician, he has a group.
It's very cool to have a big brother my sister is 10 and she goes to school. We have a
cat and a dog. I have a lot of friends, my best friend is in my class, and we are emm
we like to do something together in our free time. My friends are funny; we have fun
together, make parties, and do some crazy things. We meet with our company every
weekend. I like skating and I have a lot of friends skaters. In winter I want to
learn...to learn snowboard, but there is no snow. I like to draw and I want to learn to
draw graffiti. I'm emo, I like this style and a lot of my friends are emo too. I like
1
The square brackets here and further on mark the authentic learner text with all mistakes
preserved
A Russian EFL Learner Corpus from Scratch
173
piercing, I have one in my tongue, I want to do one in my...right here, under my lip,
but my mother doesn't like it. I like rock music and my favorite group is My
chemical Romance. I don't like pop music. Sometimes I write my own poems and
songs and stories. We often travel with my family; we sometimes go to Finland by
car. I like traveling and I dream to visit Germany - I like their language and culture.
I also like cartoons and different films and cinema. I like coffee and candies.]
6 Conclusion
Writing and speaking with timed task setting calls for a repertoire of constructions
and patterns that are simple enough and are frequently repeated in the learner output.
Their simplicity is backed by their resemblance to the mother-tongue patterns. Their
systemic preference to well-trained and more appropriate ways of expressing the same
concepts, as intermediate and upper-intermediate levels of language proficiency
demand, brings to the so-called “structural poverty”. At the same time this systemic
preference is the proof of learner’s interlanguage grammar that, very likely (since
such structures can be found in the output of non-native speakers with higher levels of
proficiency), shapes the core grammar of EFL.
Such systemic preferences are best shown in LC. The corpus data supply the
researcher with evidence that can be analyzed with different text tools. SPb LC is an
attempt to compile a target-specific structure, a text collection in accord with essential
corpus design criteria. Operated with reliable free tools, the corpus proved efficient
enough in spotting and analyzing the learner language. The SPb LC findings are
preliminary observations of learner interlanguage grammar.
References
Electronic
[1] AntConc [http://www.antlab.sci.waseda.ac.jp]
[2] AntConc 3.2.1 [http://www.antlab.sci.waseda.ac.jp/software.html]
[3] Connexor Natural Knowledge (ENGCG)
[http://www.connexor.eu/technology/machinese/demo/syntax]
[4] The Compleat Lexical Tutor for Data Driven Language Learning on the Web
[http://www.lextutor.ca/]
[5] Find and replace [ http://www.abacre.com/afr.zip]
174
Olga Kamshilova
Paper
[1] Biber D., Johansson S., Leech G., Conrad S., Finegan E. Longman Grammar
of Spoken and Written English. Harlow: Longman, 1999.
[2] Cobb T. Analyzing late interlanguage with learner corpora: Quebec replications
of three European studies // Canadian Modern Language Review, 59(3), PP.393423, February 2003.
[3] Gaskell, D., Cobb, T. Can learners use concordance feedback for writing
errors? Submitted to System, November 2003, Revision April 12, 2004
[http://www.lextutor.ca/concordancers/concord_e.html]
[4] Granger, S. Computer learner corpora, second language acquisition and
foreign language teaching. – Amsterdam: J. Benjamins Pub. Comp., 2002
[5] Granger, S., Rundell, M. From corpora to confidence // English Teaching
Professional, Issue 50, May 2007, PP.15-18 [www.etprofessional.com]
[6] Hinkel E. Simplicity without elegance: Features of sentences in L1 and L2
Academic texts // TESOL Quarterly, Vol.37, No.2, 2003, pp.275-301.
[7] Mauranen A. The Corpus of English as Lingua franca in Academic
Settings // TESOL Quarterly, Vol.37, No.3, 2003, pp.515-527
[8] Selinker, L. Interlanguage // International Review of Applied Linguistics 10, 1972,
pp. 209-31.
[9] Камшилова, О. Н., Колина, М. В., Николаева, Е. А. Разработка корпуса текстов петербургских школьников: задачи и перспективы //
Прикладная лингвистика в науке и образовании. – СПб., 2008. –
С.92-98.
[10] Разумова В. В. Выбор синтаксического анализатора для анализа
сложных предложений в представительном корпусе текстов //
Прикладная лингвистика в науке и образовании – СПб., 2008. –
СС.150-154.
Preliminary Analysis of a Slavic Parallel Corpus
Emmerich Kelih
Institut für Slawistik, University of Graz, Austria
Abstract. The focus of this paper is on a detailed description of a newlydeveloped parallel corpus of Slavic languages. It consists of 11 Slavic translations
of the well-known Russian socialist realist novel “Kak zakaljalas’ stal’/How the
steel was tempered” (KZS), written by N.A. Ostrovskij in the years 1932-34. The
KZS contains the Slovene, Croatian, Serbian (ekavian), Macedonian, Bulgarian,
Ukrainian, Belorussian, Slovak, Czech, Polish and Upper Sorbian translations.
Thus, for the first time a parallel text of almost all Slavic standard languages is
available. In addition to the discussion of some text-specific issues of KZS, an
explorative statistical analysis and a linguistic interpretation of text length and
the Type-Token Ratio is offered.
1
Introduction
Parallel texts and parallel text corpora play a crucial role in corpus linguistics, linguistic
typology and text linguistics. A parallel text is a text or part of a text placed alongside
its translation in one or many1 languages ([19, 47ff.], [17, 121ff] and [25, 73]). Parallel
corpora are explored in general linguistics and language processing. With respect to
Slavic languages in particular, the well-known “Multext East” project (cf. [6] and ([7]),
which contains many translations of George Orwell’s “1984” into Slavic languages
(Bulgarian, Croatian, Czech, Resian, Russian, Serbian and Slovene) and the ambitious
“Regensburg Parallel Corpus of Slavic Languages” [28], which includes different translated texts from and into Russian, Belorussian, Croatian, Serbian, Slovak, Czech and
Ukrainian, have to be mentioned.
According to our knowledge, however, despite the availability of these “larger”
parallel corpora focussed on Slavic languages, no parallel text corpus2 with one original
text in a large number of Slavic standard languages exists. To overcome this deficiency,
a new parallel text corpus, containing in sum 11 Slavic translations of the well-known
Russian socialist realist novel “Kak zakaljalas’ stal’/How the steel was tempered” (hereafter KZS) has been compiled by the author for a systematic cross-linguistic quantitative
analysis of Slavic languages from a synergetic and quantitative point of view ([15] [16]).
In our paper only a few selected problems can be outlined. Firstly, a short overview of
1
2
[5, 95] recently introduced the term “massively parallel texts”, defined as a huge text corpus
with translations preferably in many (genetically) diverse languages, such as, protocols of
the European Parliament in over 30 languages, translations of the “Universal Declaration of
Human Rights” and of the Bible into more than 100 languages, etc.
For more parallel-corpora projects dealing with Slavic languages cf. [28, 123] and [8], with a
project on parallel word lists of Western Slavic languages. [23] mention different available –
not solely Slavic – translations of “Le Petit Prince”.
176
Emmerich Kelih
linguistic applications of parallel texts is given; secondly, a description of the project in
progress is presented in detail; and finally the initial results of a preliminary statistical
analysis of the sample size and the Type-Token Ratio of the translations is given.
2
Parallel texts and their linguistic applications
From a linguistic point of view,3 parallel texts are an important empirical data base.
They can be used for many linguistic purposes, such as:
1. For typological and cross-linguistic analyses on the phonological, morphological,
syntactical and lexical level (cf. [11], [27], [5] and [29]); in particular, parallel texts
are a valuable basis for the study of quantitative features of language cf. ([2, 6364]).
2. For the analysis of the quality and the linguistic structure of translations (cf. [1],
[11], [12], [9], [20] and [24]).
3. For the examination of hypotheses from quantitative and synergetic linguistics. Parallel texts provide a deeper understanding of self-regulation mechanisms of translated texts; in particular, it is not known whether these special kinds of languages
abide by language laws, such as the Zipf and the Menzerath law.
4. For comparative text linguistic analyses, including the investigation of stylistic features.
5. For the study of inter-lingual readability – a research field that thus far has hardly
made use of parallel texts.
Despite this broad and appealing spectrum of various applications of parallel texts,
some remarkable arguments have been raised in the past concerning the limited linguistic and methodological usefulness of parallel texts have been raised. Sometimes it is
argued that translated texts lack “authenticity” and “quality” (cf. [22, 102], [29, 128]).
Thus some kind of “unnaturalness” of translated texts is postulated and intensively
discussed in corpus linguistics as the problem of “translationese” (cf. [18] [19, 49]).
This problem is hardly to be avoided in general, but except for this legitimate objection,
the outstanding attractiveness of parallel texts lies in the comparision of semanticallyidentical or nearly-identical texts from different languages cf. [29, 130]. Furthermore
parallel texts are at least written in a grammatically and morphologically “correct” way,
and hence they are, depending on the examined linguistic hypotheses, furthermore an
adequate and powerful empirical data base for cross-linguistic studies.
3
“How the steel was tempered (KZS)” in twelve Slavic languages
The KZS is the empirical database of an ongoing research project on quantitative phonology, namely the investigation of interrelations between the size of phoneme inventories,
3
The paper focuses on a description of some basic quantitative features of the KZS, and no
further language processing issues are discussed. For further information about sentence alignment and POS-Tagging of parallel texts cf. [27].
Exploring a Slavic Parallel Corpus
177
phoneme frequency, distribution and syllable structure in Slavic languages. In view
of the lack of available and accessible parallel texts in many Slavic languages, the
Belorussian, Ukrainian, Czech, Polish, Slovak, Upper-Sorbian4 , Bulgarian, Croatian
(ijekavian), Macedonian, Serbian (ekavian) and Slovenian translations of the novel
“Kak zakaljalas’ stal’/How the steel was tempered” (KZS) were recently collected,
scanned and submitted to OCR. All text files were then manually proofread. They are
now available as plain text.
The particular choice of this “highly” ideological text, however, is connected with
some problematic issues, which should be discussed briefly here. Firstly, the authorship
of the novel is disputed; secondly, the influence of reversions by the editors has to be
discussed; and thirdly, the base of the Russian original for the translations must be
specified.
However, today it is accepted that the novel was doubtlessly written by N.A. Ostrovskij. The final version of the novel was however “checked” by many editors, who
“polished” the text from a stylistic, and especially ideological, viewpoint ([10, 121] and
[4, 8]). In this respect – as highlighted in the detailed study by [10] – the canonical
monographic issue of KZS from 1934 differs slightly from the chapters previously
(1932-34) published separately in the Soviet literary journal Molodaja Gvardija. But
the changes in the monographic issue of 1934 in comparison with the versions published earlier are primarily related to some negligible ideological details, whereas the
macrostructure of the novel, e.g. the division into two parts, with nine chapters in each
part, remained unchanged. As a rule, the canonical text, i.e. the monograph from 1934
and it translations into the different Slavic languages, were used for the KZS. Cf. [14,
124] with a more detailled list of the scanned translations and further bibliographical
information.
Leaving aside the somewhat problematic production history, edition and translation
of the KZS, the novel – the literary and aesthetic quality can be considered low –
is a quite interesting mixture of different styles. According to [10, 140], for KZS a
simple, linear sentence structure, a high frequency of colloquial elements and in part
a “declamatory” register (‘obščestvennaja reč”) is characteristic. Additionally, a high
proportion of oral speech, mixed with a few narrative sequences and some (typical)
Soviet abbreviations and acronyms (especially in respect of the political “lexicon”)
are obtainable. Furthermore, the novel is characterised by an “internal heterogeneity”:
Throughout the novel many poems, diary entries, letters and public announcements can
be found. However, the novel is not literarily “deformed” in the strictest sense, but
rather a representative example of a literary prose text, written in a style typical for the
socialist realist writing of the 30s, in which several sub-registers of written language in
the original and translated texts can be analysed linguistically.
As already mentioned, at the macro level the KZS is divided into 18 chapters. Because of the time-consuming procedure of scanning and proofreading, only 10 chapters
of 18 (nine chapters of the first part and one chapter of the second part, with a total of
approximately 240 printed pages) of the KZS were processed for further analysis. A
4
Special attention was paid to the Sorbian languages, at least in one of the standard languages.
[26] note in their analysis of some Slavic translations of the novel Harry Potter that unfortunately no translations of this text are available for Sorbian.
178
Emmerich Kelih
more detailed explanation for this decision of a size limitation goes beyond the scope
of the present paper. For our purposes, i.e. a systematic analysis of the quantitative
structure (esp. phonological and syllable structure), the material seems to be sufficient
and should moreover be understood as case study material.
4
Quantitative characteristics of KZS
This chapter provides some promising results5 of the analysis of the text length (number
of types and tokens) and the Type-Token Ratio (hereafter TTR) of the KZS. It will
be shown that a quantitative characteristic as seemingly trivial as the sample size can
already provide more detailed information about the morphological structure of the
languages under examination. Furthermore it will be demonstrated that the TTR is an
appropriate parameter for a typological ordering of Slavic languages.
4.1
Sample size: Number of types and tokens
An important characteristic of parallel texts is the text length, measured by the number
of “words”; to be more precise in this context, by the number of tokens and types. Both
linguistic entities are here defined as orthographical units of a written text, where the
space has the function of a delimitation marker, i.e. all alphabetical signs between two
spaces are defined as a token/type. The results of the counts of tokens and types in the
KZS are summarised in Table 1 (see p. 179), where the Slavic languages are already
arranged according to their genetic/areal affiliation. Fig. 16 represents the number of
types and tokens.
The noticeable variety of the sample size7 in the parallel texts, especially in respect
of the number of tokens, can probably be explained by different morphological and
syntactical characteristics of the Slavic languages. To start with the Eastern Slavic
languages (Russian, Ukrainian, Belorussian), it appears that they share approximately
the same text size with respect to the number of types and tokens. The difference in
relation to Russian of 63 tokens (Ukrainian) and 335 tokens (Belorussian) is relatively
5
6
7
For a comparative analysis of KZS grapheme frequencies cf. [13].
Slo = Slovene, Cro = Croatian, Serb = Serbian, Bulg = Bulgarian, Mz = Macedonian, Sorb
= Upper Sorbian, Rus = Russian, Ukr = Ukrainian, Belorus = Belorussian, Cz = Czech, Sk =
Slovak and Pol = Polish
The parallel texts have not yet been annotated and tagged. For Russian, Serbian and Slovene a
sentence alignment (with the help of Duško Vitas, Belgrade) has already been performed. The
author is grateful for any cooperation with specialists of tagging, annotation and lemmatisation
of Slavic texts. Even if the KZS has thus far not been processed adequately, some quantitative
studies can at least be performed. It can also be claimed that all texts bear approximately equal
semantic information, and that all texts do not differ in terms of the original version used for
the translations. On the one hand all scanning and proofreading has been done manually and on
the other hand the text length of all ten chapters of each language was statistically correlated
with the chapter text length of the Russian original. The interrelation of the text length in
translated and original texts can be described by simple linear models, with a sufficient R2 >
0.98 in all examined cases. Thus, it is likley to conclude that for all translated texts the same
source was used and there are no significant stylistic modifications in the translations.
Exploring a Slavic Parallel Corpus
›
1
2
3
4
5
6
7
8
9
10
11
12
Language group
South Slavic
Eastern Slavic
Western Slavic
Language
Tokens
Types
TTR
Slovene
Serbian
Croatian
Bulgarian
Macedonian
Russian
Ukrainian
Belorussian
Czech
Slovak
Polish
Sorbian
62655
56230
56424
57174
58837
49675
49612
50010
52180
52099
52737
58484
13946
13642
13737
12308
11465
15053
14645
14858
14136
14027
14978
14574
4.4927
4.1218
4.1074
4.6453
5.1319
3.3000
3.3876
3.3659
3.6913
3.7142
3.521
4.0129
179
Table 1. Text length (types, tokens) and TTR of KZS
marginal, so it can be concluded that the Eastern Slavic languages do not show notable
differences between the text lengths, either at the tokens or at the types level.
A relativley similar picture is obtained for the Western Slavic languages: The Slovak, Polish and Czech translations have a more or less similar text length (approx. 52300
tokens). This especially holds true for Slovak and Czech, with a difference of just 82
tokens. The text length of Polish is somewhat higher, for instance in relation to Czech,
which has 550 tokens more. A clear outlier within the Western Slavic languages is the
Sorbian translation, which has 6304 tokens more than the Czech text. This difference
can perhaps be explained by morphological differences between Czech and Sorbian,
such as the analytic form of the past tense, the formal expression of the reflexivity,
etc. This is a hypothesis which of course must be studied in more detail in the future.
However it is worth mentioning that on the types level the difference is not so for the
Sorbian text, which has only 438 types more than the Czech text.
Further evidence for our claim that morphological characteristics are responsible for
the differences in text length can be found in South Slavic languages. The genetically
“very close” languages of Croatian (56424 tokens) and Serbian (56230 tokens) do not
show any notable differences at all; similarly small differences regarding the number
of types and tokens can be found for the Bulgarian and Macedonian texts. Again, just
one language, namely Slovene, demonstrates a behaviour that is in some way specific
regarding the number of tokens. Slovene, in comparision with other Slavic languages,
shows a relativley “normal behaviour” on the types level, on the tokens level it is one of
the longest texts (62655). This striking feature – at least on the tokens level in relation to
the Russian text (49675 tokens) – can again be explained linguistically: In Slovene, the
analytical form of the past tense by means of the auxiliary verb “to be/je”, the intensive
usage of the analytical past perfect (“je bil vedel/be was known”) and the frequent use
of relative clauses are relatively typical and commonly used. Russian neither makes use
of auxiliary verbs nor has a complex analytical past tense form and thus for the Slovene
text a frequent use of synsemantic words and a high number of tokens can be observed.
180
Emmerich Kelih
Fig. 1. Types and tokens in Slavic parallel texts
All in all, a fairly simple analysis of the text length already provides some in-depth
information about the morphological and syntactical structure of the parallel text corpus
under examination.
4.2
Type-Token Ratio
In addition to a simple description and interpretation of the number of types and tokens
of the parallel texts, a more detailed examination of the so-called Type-Token Ratio
(TTR = Tokens/Types) is required. Whereas in the past this index was primarily understood as an indicator of lexical richness (cf. the overview on possible interpretations of
the TTR in [3, 108]), the TTR – especially in comparative studies of parallel texts –
can be introduced as a measurement of the morphological richness of word forms and
the productivity of the flexion system. The more word form types a text (=language)
has, the greater is the variety of morphological forms that it contains and thus a low
TTR in parallel-text research can be interpreted as an indicator of the synthetism of
one language. For more in-depth research of word frequencies in language typology cf.
[21]. The TTR for all examined Slavic languages is graphically represented in Fig. 2
and already ranked from its maximum to its minimum.
It can clearly be seen from Fig. 2. that, based on the TTR, a typological order of
the Slavic languages can be obtained: It starts with the strongly analytic Southeast
Slavic languages, Macedonian and Bulgarian, and continues with Slovene, Serbian and
Croatian. They are followed by the Western Slavic languages (order of languages: Sor-
Exploring a Slavic Parallel Corpus
181
Fig. 2. TTR in Slavic parallel texts
bian, Slovak, Czech and Polish), which appear as one “group”8 . Finally, the East Slavic
languages (Ukrainian, Belorussian and Russian) have the lowest TTR and also appear
as one typological “row”. In sum, the TTR of parallel texts can be interpreted alongside
the morphological richness as an indicator of the degree of analytism/synthetism of the
languages.
5
Conclusion
The paper has the main function of presenting the KZS parallel-text corpus, which
should, of course, be refined in the future, especially from the viewpoint of language
processing (tagging, alignment, and lemmatisation). As shown in our exploratory discussion, the text size (number of types and tokens) and the Type-Token Ratio of parallel
texts appear to be a simple yet efficient tool for obtaining language-specific morphological behaviour of the parallel texts analysed. Moreover, strong evidence for the general
usefulness of parallel texts for language typology is given.
References
[1] Altenberg, B. and Aijmer, K. (2000). The English-Swedish Parallel Corpus: A
resource for contrastive research and translation studies. In Ch. Mair and Chr.
8
In future more rigorous methods such as cluster methods must be applied.
182
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
Emmerich Kelih
Hundt, editor, Corpus linguistics and linguistic theory. Papers from the Twentieth
International Conference on English Language Research on Computerized Corpora (ICAME 20), Freiburg im Breisgau 1999, pages 15–33, Amsterdam. Rodopi.
Altmann, G. and Lehfeldt, W. (1973). Allgemeine Sprachtypologie. Fink,
München.
Altmann, V. and Altmann, G. (2008). Anleitungen zu quantitativen Textanalysen.
Methoden und Anwendungen. RAM-Verlag, Lüdenscheid.
Anninskij, L. (1989). Obručennyj s ideej. In Ostrovskij, N. A., editor, Sobranie
sočinenij v trech tomach. Tom 1. Kak zakaljalas’ stal’, pages 7–28, Moskva. Molodaja Gvardija.
Cysouw, M. and Wälchli, B. (2007). Parallel texts: using translational equivalents
in linguistic typology. Sprachypologie und Universalienforschung, 60(2):95–99.
Dimitrova, L., Ide, N., Petkevič, V., Erjavec, T., and Tufis, , D. (1998). MultextEast: Parallel and Comparable Corpora and Lexicons and Lexicons for six Central
and Eastern European Languages. In Proceedings of the 36th annual meeting on
Association for Computational Linguistics. Volume 1, pages 315–319, Montreal
and Quebec.
Erjavec, T., Ide, N., Petkevič, V., and Véronis, E. (1995). Multext-East: Multilingual Text Tools and Corpora for Central and Eastern European Languages. In
Language Resources for Language Technology: Proceedings of the TELRI (TransEuropean Language Resources Infrastructure) European Seminar (1st, Tihany,
Hungary, September 15-16, 1995), pages 88–97.
Garabík, R. and et al. (2007). A Cross-linguistic Database of Children’s Printed
Words in Three Slavic Languages. In Levická, J. and Garabík, R., editors, Slovko
2007. Fourth International Seminar. Bratislava, Slovakia, 25–27 October 2007,
pages 51–64, Bratislava. Tribun.
Gellerstamm, M. (1996). Translations as a source for cross-linguistic studies.
In Aijmer, K., Altenberg, B., and Johansson, S., editors, Language in Contrast:
Papers from a symposium on Text based Cross-linguistic studies. Lund, March
1995, pages 53–62, Lund. Lund University Press.
Guski, A. (1981). N. Ostrovskij: Kak zakaljalas’ stal’: biographisches Dokument
oder sozial-realistisches Romanepos? Zeitschrift für slavische Philologie, 42:116–
145.
Johansson, S. (1998). On the role of corpora in cross-linguistic research. In
Johansson, S. and Oksefell, S., editors, Corpora and Cross-Linguistics Research,
pages 3–24, Amsterdam. Rodopi.
Johansson, S. (2003). Reflections on Corpora and their Uses in Cross-linguistic
research. In Zanettin, F., Bernardini, S., and Stewart, D., editors, Corpora in
translator education, pages 135–144, Manchester. St. Jerome Publisher.
Kelih, E. (2009a). Graphemhäufigkeiten in slawischen Sprachen: Stetige Modelle.
Glottometrics, 18:53–96.
Kelih, E. (2009b). Slawisches Parellel-Textkorpus: Projektvorstellung von “Kak
zakaljalas’ stal’ (KZS)”. In E. Kelih, E., Levickij, V. V., and Altmann, G., editors,
Methods in text analysis, pages 106–124, Černivci. Ruta.
Köhler, R. (1986). Zur linguistischen Synergetik: Struktur und Dynamik der Lexik.
Brockmeyer, Bochum.
Exploring a Slavic Parallel Corpus
183
[16] Köhler, R. (2005). Synergetic linguistics. In Köhler, R., Altmann, G., and
Piotrowski, R. G., editors, Quantitative Linguistik/Quantitative Linguistics. Ein
internationales Handbuch. An International Handbook, pages 760–774, Berlin.
de Gruyter.
[17] Lemnitzer, L. and Zinsmeister, H. (2006). Korpuslinguistik: eine Einführung.
Narr, Tübingen.
[18] Mauranen, A. (2002). Will ‘translationese’ ruin a contrastive study? Languages
in Contrast, 2:161–186.
[19] McEnery, T., Xiao, R., and Tono, Y. (2006). Corpus-based language studies: an
advanced resource book. Routledge, London.
[20] Mohanty, P. (2008). The Semantic Differential Technique and Measurement of
Translational Meaning. In Altmann, G., Zadorozhna, I., and Matskulyak, J., editors, Problems of General, Germanic and Slavic Linguistics. Papers for the 70th
anniversary of Professor V.V. Levickij., pages 215–225, Černivtsi. Knichi XXI.
[21] Popescu, I.-I. and Altmann, G. (2008). Hapax Legomena and Language Typology.
Journal of Quantitative Linguistics, 15(4):370–378.
[22] Stolz, T. (2007). Harry Potter meets Le petit prince: On the usefulness of parallel corpora in crosslinguistic investigations. Sprachtypologie und Universalienforschung, 60(2):100–117.
[23] Stolz, T., Stroh, C., and Urdze, A. (2007). Nicht ganz ohne [. . . ]. In Grzybek,
P. and Köhler, R., editors, Exact Methods in the Study of Language and Text.
Dedicated to Professor Gabriel Altmann on the Occasion of His 75th Birthday.,
pages 633–646, Berlin and New York. de Gruyter.
[24] Teubert, W. (2002). The role of parallel corpora in translation and multilingual
lexicography. In Altenberg, B. and Granger, S., editors, Lexis in Contrast, pages
189–214, Amsterdam. Benjamins.
[25] Teubert, W. and Čermáková, A. (2007). Corpus linguistics: A short introduction.
Continuum, London.
[26] van der Auwera, J., Schallea, E., and Nuyts, J. (2005). Epistemic possibility in a
Slavonic parallel corpus – a pilot study. In Hansen, B. and Karlík, P., editors,
Modality in Slavonic languages. New perspectives, pages 201–217, München.
Sagner.
[27] Véronis, J. (2000). From the Rosetta Stone to the Information Society: A Survey of Parallel Text Processing. In J. Véronis, editor, Parallel Text Processing.
Alignment and Use of Translation Corpora, pages 1–25, Dordrecht. Kluwer.
[28] von Waldenfels, R. (2006). Compiling a Parallel Corpus of Slavic Languages. Text
strategies, Tools and the Question of Lemmatization in Alignment. In B. Brehmer
and V. Zhdanova and R. Zimny, editor, Beiträge der Europäischen Slavistischen
Linguistik (POLYSLAV) 9, pages 123–138, München. Sagner.
[29] Wälchli, B. (2007). Advantages and disadvantages of using parallel texts in typological investigations. Sprachypologie und Universalienforschung, 60(2):118–
134.
Operators for Extending and Developing an Utterance
(Based on Operators of Concessive Relation)
Jana Kesselová
University of Prešov in Prešov, Slovakia
Abstract. The paper deals with operators of concessive relation which we
consider to be a transitional type between operators used for developing an
utterance and operators used for extending an utterance. An empirical source
of the research is the Slovak National Corpus. The aim of this research is to
enrich the existing paradigmatic approach to classes of operators by a syntagmatic approach focused on their real textual “behaviour”. We proceed from
the form of an operator through its contextual setting towards more general
aspects of its functioning. At the same time we concentrate on the inventory
of concessive operators in contemporary written Slovak, types of relations
between utterance meanings, and communication functions of utterances with
individual types of operators. Contemporary written Slovak has an extraordinarily miscellaneous and formally, stylistically, distributionally and
pragmatically differentiated inventory of concessive operators. Dynamics in
this microsystem is demonstrated by strengthening the position of a prototype
representative and identifying the centre with synchronically most frequent
operators on the one hand and decline (even disappearance) of some synonymous operators on the other hand. Another example of dynamics is stylistic
and pragmatic specification of operators.
1 Introduction
While basic word classes have been described and analysed in Slovak linguistics at
least in one monograph each, non-basic word classes do not belong to the centre of
linguists' research attention (one of exceptions is the period of preparation of a book
Morfológia slovenského jazyka, 1966). Nevertheless, when forming an utterance and
reaching a communicative intention, means of joining isolated naming units into
coherent textual sequences are as important as the names of independent objects and
attributes.
Relations between phenomena present a universal system of connections and
associations or, more precisely, people interpret them as a system of mutual
connections and associations and they process them linguistically in communication
[5]. Among means oriented functionally towards establishing a system of relations,
the article focuses on connecting expressions and their functioning in a real
language interaction. The centre of connecting expressions consists of words with a
primarily connecting function, i.e. conjunctions and their combinations. Other
connecting means are e.g. interrogative pronouns (their connecting function is only
secondary), combinations of conjunctions and pronouns, and fixed phrases
consisting of conjunctions and other word classes. Thus instead of using the term
conjunctions I prefer the term with a much broader sense – operators. Operators for
extending and developing an utterance are understood as language means with a
Operators for Extending and Developing an Utterance
185
connecting function which join clause elements and clauses into higher utterance
and textual units. They are formally miscellaneous and their centre consists of words
with a primarily connecting function, i.e. conjunctions. These are supplemented by
further connecting means:
• relatives (interrogative pronouns in which the connecting function is realized
only as secondary);
• co-relatives (co-relative pairs of conjunctions and demonstrative pronouns or
interrogative and demonstrative pronouns);
• conjunctional expressions (combinations of conjunctions and other word
classes, mainly particles, primary and secondary prepositions and adverbs).
The second reason why I prefer using the term operators to the term conjunctions
is motivated by a communicative-functional approach to the role of connecting
means in human communication. From the communicational point of view,
operators do not have only a connecting function, but they take part in creation or
modification of the communication function of the utterance which is understood as
a purpose (intention) of communication. Thus I concentrate on operators not only as
means of establishing links and relations between utterance meanings, but also as
indicators and constituents of communication functions of utterances.
Distribution of operators used for extending and developing an utterance is based
on the fact that the “lower” units enter various relations in higher utterance units and
they more or less lose their autonomy. According to the relation which operators
establish between utterance meanings, two types of operators may be distinguished
[3]:
• Related utterance meanings preserve their high degree of autonomy and the
relation of relative semantic independence. Operators present one utterance meaning
in relation to another one as simply coordinated, possible, contradictory or more
important. Operators in utterances with semantically independent utterance
meanings are considered to be operators for extending the utterance.
• When creating an utterance, utterance meanings may lose their autonomy
because they enter the relation of semantic dependence. It happens in the case when
one utterance meaning is determined by another one (e.g. if some action/state is
presented as a cause, purpose, condition, temporal circumstance, etc.). Operators in
utterances with semantically dependent utterance meanings are considered to be
operators for developing the utterance.
The study focuses on operators of concessive relation which I consider to be a
transitional type between the above-mentioned types.
2 Method of research
An empirical source of the research is a large number of written texts in the Slovak
National Corpus. The research is based on a so-called balanced corpus, in which
publicistic texts (33,3%; doc.type=inf), scientific texts (33,3%; doc.type=prf) and
belletristic texts (33,3%; doc.type=img) are present. I have used data about the total
frequency of a phenomenon in the Corpus which is marked as its synchronic
186
Jana Kesselová
communication loading in a particular corpus [17]; about the stylistic value of an
operator (concordances → statistics → attribute doc. type → frequency
distributions); and about the most frequent collocations (concordances → statistics
→ attribute: word/lema → most frequent collocations). To avoid including
homonymous expressions into the total number of occurrences of individual
operators, I have separated operators from other word classes with a command, e.g.
[(word="hoci") & (tag="O.*")] or, where necessary, also with a negative filter or
manually.
Thanks to the existence of the Corpus, it is possible to complete, specify and
refine the existing conclusions (based and verified mainly on belletristic texts) and
enrich the existing paradigmatic approach to classes of operators with the
syntagmatic approach aimed at their real textual “behaviour”. I have proceeded from
the form of operator through its contextual setting towards more general aspects of
its functioning. At the same time I have concentrated on:
• the inventory of concessive operators in contemporary written Slovak
The system of operators provides a set of possibilities, an offer for a producer of
the text. I am interested in a) motivation of selection of a particular operator and b)
competing and complementary relations between central and peripheral operators.
• types of links and associations between utterance meanings
Meaning associations and links which are created by operators have been
investigated in a selective set of utterances with a given operator. With regard to a
high frequency of operators, it is beyond possibilities of an individual to investigate
all the utterances. It is known that one of methodologic questions of corpus
linguistics is representativeness of a sample created from the “profusion” of
occurrences which the corpus offers. When creating a representative set, I have
taken into consideration two criteria. The criterion of contextual surroundings means
that utterances with the most frequent collocations of the operator have been
included in the set (surprising semantic homogeneity of expressions occurring in the
surroundings of operators has been proved). The second criterion is a stratified way
of utterance selection into the research sample. It means that utterances from
publicistic, scientific and belletristic texts have been included in the sample equally
(if the calculation of frequency distribution has shown stylistic neutrality of the
operator) or unequally, according to the calculated value of operator frequency
distribution.
• communication functions of utterances with individual types of operators
In complex books on Slovak morphology and syntax [9, 10] conjunctions are
characterised either as a set of means serving for expressing a certain relation, or as
enumeration of relations corresponding with individual conjunctions. The aim of
this corpus research is to show that operators are not only means used for
establishing relations, but that they together with lexical surroundings, functional
sentence perspective, etc. participate in creating communication functions of the
utterance.
Operators for Extending and Developing an Utterance
187
3 Concessive relation and its character
A concessive relation is traditionally characterised within the frame of dependent
adverbial relations in the Slovak linguistics [10, 11] with the explanation that
dependent adverbial clauses of concession express circumstances unfavourable to
the action of the main (independent) clause.
On the one hand, I consider concessive relations to be a part of causal relations
in a broad sense of word. Slovník slovenského jazyka III [14] defines concession as
a real or possible cause whose expected consequence (effect) in a given instance
does not come. The basis of the concessive relation is to realise the causality
between two phenomena which is regularly or usually realized (x is always/usually
the cause of y). The concessive relation is a specific aspect of causal relation in that
sense, that it includes instances which go beyond regularity or commonness. In the
relation cause – consequence it “counts” with the influence of unexpected
circumstances, e.g.:
(1) Hoci má Irak ohromné zásoby ropy, nemá benzín.
The empirical context (knowledge that petroleum is used for petrol production)
leads to the expectation that petroleum resources in Iraq mean also sufficiency of
petrol reserves. However, the utterance with the concessive operator hoci brings
unpresupposed (on the basis of the presupposition “who has petroleum has also
petrol”) even surprising conclusion (favourable circumstance x brings unfavourable
situation y). In other words, operators of concessive relations construct utterances
which demonstrate linguistically the fact that the relation between cause and its
consequence has only relative validity. P. Karlík [5] defines a concessive relation as
a relation between ineffective cause and unexpected consequence.
On the other hand, the concessive relation is very close to a contrast relation
since its utterance meanings are in the mutual relation of disagreement even
contrast. This “semantic nearness” is reflected either in the use of combinations of
concessive and contrast operators (hoci – ale, hoci – no, hoci – predsa, hoci –
jednako)or in a possible mutual substitution of concessive and contrast operators:
(2a) Hoc si moji exspolužiaci vybavujú hypotéky a kupujú byty, mne stále vyhovuje
bývanie v prenájme.
(2b) Moji exspolužiaci si vybavujú hypotéky a kupujú byty, ale mne stále vyhovuje
bývanie v prenájme.
However, it is necessary to add that the potential substitution of operators relates
to non-contextual utterances; substitution of operators in the context modifies the
communication function of the utterance. As for their function, concessive operators
are very close to causal and contrast operators. Thus I consider them to be a
transitional type between operators for developing and extending the utterance.
188
Jana Kesselová
4 Concessive operators
4.1 Inventory of operators and their position in the utterance
The inventory of concessive operators is a very diverse system which is formally,
stylistically and pragmatically differentiated. From the formal point of view it
consists of the conjunctions hoci (and its variant hoc) and síce, and the
conjunctional expressions which contain basic subordinating conjunctions and
relatives:
• keď (aj keď, i keď, ani keď, keď ešte);
• keby (aj keby, keby aj, i keby);
• aby (namiesto toho, aby);
• že (napriek tomu, že; navzdory tomu, že; bez ohľadu na to, že; nehľadiac na to,
že; nezávisle na tom/od toho, že);
• čo (a čo aj, a čo i, čo hneď, čo ako, čo priam);
• ako (akokoľvek).
Each of them may enter a binary conjunctional expression, e.g.: hoci – predsa,
hoci – jednako, i keď – aj tak, etc. Moreover, all operators have both, initial and
non-initial positions in the utterance. The non-initial position strongly prevails in
communication(the only exception is the conjunctional expression nehľadiac na to,
že, which occurs equally in both positions). Thus the order of clauses is not
motivated only by the principle of constructional iconism (the order cause –
consequence in reality), but also by the communication intention of a speaker/writer
(which part of the utterance meaning he/she wants to emphasize – either ineffective
cause or unexpected consequence), e.g.:
(3a) Ružomberská trénerka môže byť so začiatkom série spokojná,, hoci k niektorým
fázam zápasov mala určité výhrady.
(3b) Hoci dosiaľ obce návrhy nemuseli zverejňovať, viaceré to už robili aj predtým.
A different position of the operator in the utteranceis a stylistic indicator (for
example, the non-initial position of the operators hoci, aj keď, i keď is typical of
belles-lettres style while the initial position is typical of publicistic texts). The
Corpus enables us to find out even more minute differences. Different positions of
operators in utterances are bound with different collocations which leads to
modifications in communicative functions of utterances. The operators i keby, aj
keby, when being in the initial position, co-create utterances in which the speaker
admits a very improbable, even unreal circumstance in favour of ensuring the
addressee about the truth value of his/her claim (Aj keby som mal skončiť na
šibenici..., Aj keby som mal zomrieť dvakrát..., Aj keby som mal sto rokov sedieť
v base..., Aj keby som mal do konca života kopať kanály..., I keby som mal
k dispozícii tisíc ľudí..., I keby som mal sto rúk....).
Operators for Extending and Developing an Utterance
189
(4a) I keby som mal donútiť rieky, aby tiekli proti prúdu, a slnko, aby nesvietilo, aj
tak splním tvoje želanie.
On the other hand, the operators i keby, aj keby in the non-initial position
participate in utterances whose communication function is an urgent appeal to take
action even in the case of unfavourable consequences.
(4b) Ochraňujte vaše kone, zastavte ich vývoz, aj keby to malo byť na úkor vašich
platov.
The most frequent collocations show that unfavourable circumstances are
described in an impersonal way, e.g.: aj keby to malo znamenať vojnu, aj keby to
malo znamenať smrť, aj keby to malo znamenať zvýšenie daní, aj keby to malo ísť
dolu vodou, aj keby to malo vyznieť v náš neprospech, aj keby to malo trvať sto
rokov, aj keby to malo mať podobu horšieho, aj keby to malo znamenať, že sa stratí
čas, etc. Impersonal constructions preserve the principle of utterance quality, but at
the same time they express unfavourable consequences more softly than directly
addressed constructions, compare e.g.: aj keby si mal zomrieť – aj keby to malo
znamenať smrť.
4.2 Synchronic communication loading and the stylistic value of concessive
operators
Total occurrence of an operator in the Corpus reflects its synchronic communication
loading. Frequency of various operators differs remarkably in communication (see
diagram 1). According to J. Oravec [10] concessive conjunctions consist mainly of
two bases – hoci, čo. The corpus research has shown that the conjunction hoci and
operators with the component čo represent two opposite poles of frequency loading
of concessive operators: while the conjunction hoci is dominant among concessive
operators and it represents nearly 40 % of their uses, the operators with the
component čo are on the decline (and are felt as literary).
(5a) S úradmi sa nepri, syn môj, hovorievali mu otec, a čo aj neboli najlepším
rodičom, veru mali pravdu.
(5b) Tento zámok nedobyjeme, čo hneď pod ním vycedíme ostatnú kvapku krvi.
(5c) Báseň je apoteózou slobody: „...a čo i tam dušu dáš v tom boji divokom...
(5d) Nuž čo priam Ľudmilôčka moja nič nedostane, veď si ju ja vyživím.
190
Jana Kesselová
Diagram 1. Synchronic frequency loading of concessive operators
The four most frequent operators – hoci, síce, aj keď, napriek tomu, že cover
90 % of concessive operator uses. The operators hoci, aj keď are used mainly in
belletristic texts, the operators hoci, síce, aj keď, napriek tomu, že; napriek tomu, že
are used mostly in publicistic texts (the operators used in the initial position have a
capital letter at the beginning). Similarly, other, less frequent concessive operators
are used mostly in the belletristic or publicistic style; the only conjunction hoci
prevails mainly in the scientific style.
In my opinion, stylistic distribution of operators is the outer reflection of the
character of concessive relation (part 3). It is known that a scientific text is based
mostly on the explanatory composition style whose basic feature is explicativeness.
According to J. Dolník and E. Bajzíková [1] explication relates to statements about
facts which are given; where there is no doubt, but question leading towards
understanding their existence. On the other hand, concessive operators establish a
relationship between contradictory utterance meanings and form contradictory
statements and unexpected or surprising associations and connections. This raises
doubts about presented facts and evokes the expectation that the facts are to be
justified properly. Compare:
(6a) Keďže nič neje, nepriberá.
(6b) Hoci nič neje, priberá.
In the first utterance (6a) the causal operator establishes a relationship between
cause and consequence in the form of a statement; in the second utterance (6b) the
concessive operator introduces a contradictory statement which raises doubts and
also expectation that untypical consequence will be explained or justified. The
relation between causal and concessive operators resembles the relation between
explicationand argumentation to the intent that causal operators are connected with
the explicatory expansion of a topic and concessive operators are more connected
Operators for Extending and Developing an Utterance
191
with the argumentation expansion of the topic. Argumentation can be either analytic
or practical. According to J. Dolník and E. Bajzíková [1] practical (non-analytic)
argumentation is used in common (non-scientific) communication and analytic
argumentation is used in scientific texts.
Utterances with concessive operators in the analytic argumentation have a form
which meets the requirements for the quality of utterance understood to the intent
that we should not say anything without having sufficient evidence [15]. Concessive
operators establish relationships between utterance meanings not only to emphasize
various points of view, but they even stimulate recipient's multi-aspectual point of
view and thus they raise possible objections, look for evidence,and finally find the
truth. Compare:
(7a) Hoci Šimonovičove preklady majú svoje prednosti, medzi ktorými vyzdvihujem
najmä priekopnícku prácu s asonanciou, resp. s eufonicky nepresným rýmom, a hoci
jeho preklady radu básní všestranne obstoja v silnej konkurencii a možno ich pokladať za adekvátne, predsa len funkciu základného prekladu Lorcovej poézie ako
relatívneho celku aj pre slovenského čitateľa plnia skôr...)
(7b) Keby aj Oldřich Kulhánek, absolvent vysokej školy umeleckopriemyselnej
u profesora Svolinského, nebol autorom celého emisného programu platidiel Českej
národnej banky, jeho grafické dielo pre svoje kresebné majstrovstvo už i tak vošlo
do pokladnice českej výtvarnej histórie, rozsahom námetov i hĺbkou filozofickej
výpovede.
Concessive operators also constitute utterances of practical argumentation which
are not about searching for the truth in a scientific sense of the word. According to J.
Dolník and E. Bajzíková [1] the practical argumentation is used to justify
propositions which are beyond proving in a strict logical sense. It is not about
proving the truth value of the proposition, but about showing its validity. It can be
used for justifying certain behaviour or attitudes in everyday life.
8a) Uč sa pozorovať a chápať jej situáciu, hoci má chyby, predsa môžeš čerpať z jej
skúseností.
(8b) Aj rodičia boli donútení k tomu, aby dali výpoveď, hoci inú robotu ešte nemajú.
(8c) No hoci do Škandinávie odcestoval, predsa tam nezostal a po poslednej previerke vo Švédsku sa vrátil domov.
(8d) Obedy mu prihrieva stará pani zo susedného bytu, prilipol k nej, hoci s cudzími
sa zbližuje len ťažko.
In both types of argumentation it is possible to evaluate the communication
function of utterances with concessive operators as justifying or giving reasons (an
effort to show or prove validity of a proposition) or as ensuring the addressee about
the truth value of the content which seems to be improbable, unpleasant or contrary
to the existing experience. In my opinion, concessive operators belong to traditional
ensuring means such as, for example, explicit performative formulae, e.g. ubezpečujem vás or modality particles.
192
Jana Kesselová
4.3 Communication functions of concessive operators
Preference of some of competing concessive operators is motivated by a
communication intention of a speaker. Utterances with concessive operators contain
a statement and also correction of a conclusion which is made by a recipient (on the
basis of the given statement). Thus the utterance meaning joined by the concessive
operator restricts, relativizes or more or less denies the truth value of the
proposition.
(9a) Podarilo sa im urobiť vynikajúci album plný skvelých piesní, aj keď nie sú
celkom originálne.
(9b) Formálne rozhovory na túto tému neprebehli, hoci priznáva, že sa o tom v kuloároch rozprávalo.
(9c) Doteraz to bolo celkom idylické, i keď nie celkom neproblematické.
With regard to a large number of examples found in the Corpus (the number of
utterances with concessive operators presented in this paper is more than 460
thousand in the Corpus prim-4.0-public-all), I have worked with selected sets
created according to two criteria (most frequent collocations and stratified selection
based on stylistic characteristics). Results can be characterised as tendencies which
reveal a large number of real syntagmatic realizations but they definitely “resist”
any rigorous classification criteria. Beneath I present some of the results.
4.3.1 Right-hand collocations of the four most frequent operators (hoci, síce, aj keď,
napriek tomu, že) show semantic preference and semantic proximity of expressions
with which operators combine in utterances. The most frequent collocations of
dominant concessive operators refer to a variance between:
• how phenomena and events appear/are presented and how they really are
(according to the speaker), e.g.: navonok, naoko, zdanlivo, zdať sa, oficiálne,
oficiálny, formálne, deklarovať, proklamovať
(10a) Bol odjakživa mojím skrytým nepriateľom, hoci navonok sa mi zaliečal viac,
ako by človek čakal.
• prevailing or generally accepted opinion and speaker's opinion (väčšina,
mnohí, viacerí, všeobecne, niektorí, podaktorí, väčšinovo)
(10b) Myslím si, že sa nám podarilo udržať určitú vážnosť architektúry v spoločnosti, hoci väčšina architektov to stále vidí tak, že nie.
• situation or conditions in the past and at present (pôvodne, spočiatku, vlani,
pôvodne, predtým)
(10c) Nelamentuje, ani keď z nej ťahám peniaze, hoci pôvodne predpokladaná suma
na stužkovú sa zdvojnásobila.
Operators for Extending and Developing an Utterance
193
A common factor of such constituted utterances is their non-operative function
which oscillates between statement, evaluation and critique. O. Müllerová [8]
characterizes operativeness as intensity of speaker´s influence on his/her communication partner. The aim of operational utterances is to influence partner´s
behaviour; in non-operational utterances the speaker focuses more on expressing
himself/herself.
Except for common tendencies, it is possible to observe also specific textual
“behaviour” in syntagmatic relations. Typical surroundings of the conjunction hoci
are verbs like uznávať, priznávať, pripúšťať, netvrdiť, nevylučovať, nepochybovať,
nepopierať. The most frequent collocations of the conjunction aj keď are except for
the above-mentioned verbs also quantifiers, e.g.: aj keď + nie vždy, nie všetok, nie
celkom, nie veľmi, nie úplne, nie príliš, nie každý, nie všade, nie celý. The result of
these combinations are utterances in which the speaker presents his/her opinion, but
at the same time he/she“leaves the back door open”:
(11a) Už dávno sa nepovažuje za ľavičiara, hoci nepopiera, že vyšiel z tohto
prostredia.
(11b) Naozaj sme hrali postupne o čosi lepšie, hoci netvrdím, že to bola prevratná
zmena.
(11c) Šancu máme, aj keď netvrdím, že veľkú.
(11d) Tanečníci majú potenciál, aj keď nie všetci ho využijú.
A similar communicative function, i.e. to express opinion but in a relativising
way may be observed in utterances with the operators síce and napriek tomu, že in
combinations with the most frequent collocations. Here belong:
• positive evaluative expressions (síce + fajn, chvályhodný, lákavý, efektný,
snaživý; napriek tomu, že + obsiahly, skutočný, veľmi účinný, atraktívny, dlhodobý,
náročný, jednoduchý, najväčší, výhodný);
• verbs in a negative form (síce + neuviesť, nespochybňovať, neodstrániť,
nenahradiť, nezakazovať, nevyznačovať sa, nevyhlasovať, nepredpisovať, neudeľovať;
napriek tomu, že + nesprávať sa, nehrať, nepoznať, neexistovať, nezískať, síce +
nevynikať, neobhájiť, neskórovať, neumiestniť sa, neoslniť, nebodovať,
nezvíťaziť/prehrať, nepredviesť).
These combinations of expressions create utterances in which the speaker admits
quality of a certain phenomenon but relativizes it immediately by presenting this
phenomenon from a different angle. Such types of utterances may express multiaspectual, non-simplified or untraditional depiction of things and events (12a),
critique (12b) or defence reaction to critique. The speaker admits certain negatives,
but immediately counterbalances them with positives (12c, d). The operators síce
and napriek tomu, že occur mostly in publicistic texts.
194
Jana Kesselová
(12a) Snaha dostať ženy do politiky je síce chvályhodná, riešiť to však prostredníctvom zákona nie je najšťastnejšie.
(12b) Príslušník mladšej generácie sa v antológii nenájde, a to aj napriek tomu, že
ide o výber naozaj obsiahly, takmer štyristostranový.
(12c) Kňazovický síce neobhájil zlato na neolympijských 200 m, ale ani bronz nie je
neúspechom.
(12d) Zmeny v zákone síce neodstránili všetky nedostatky, ale určite eliminovali tie
najväčšie.
4.3.2 The producer of the text often takes an emotionally evaluative attitude to an
unexpected consequence. The unexpected consequence may cause a pleasant
surprise, astonishment and positive attitude. This is usually expressed explicitly
(13a, b) or it follows from the utterance content. A specific place belongs to the conjunctional expression keď ešte which constitutes utterances with the communication
function of surprise in sports and economic publicistics (13c, d).
(13a) Hoci koncertné albumy sa zvyčajne vydávajú po niekoľkých radových
platniach, Portishead sa k tomu odhodlali už po dvoch. Bol som tým veľmi milo
prekvapený... (13b) Som rád, že fanúšikovia si všímajú mňa namiesto toho, aby
povzbudzovali svoje mužstvo.
(13c) Bardejovčania v minulej sezóne vyhrali v Nitre 4:3, keď ešte v 85. minúte
prehrávali 2:3.
(13d) Česká koruna oslabila na hodnotu 38,530 CZK/1EUR, keď ešte počas
ranného obchodovania atakovala úroveň 38,350/1EUR.
However, disappointment and disillusionment in the analysed utterances prevail
overpleasant surprise from the unexpected consequence. This emotional-evaluative
attitude is indicated mainly by the concessive operators čo ako – akokoľvek – ani
keď – i keby/aj keby which are used mostly in belletristic texts. Each of them has a
very specific meaning and collocations, but their common communicative function
may be expressed with this relation: hoci + a negative emotional attitude of the text
producer.
The conjunctional expresion čo ako is preferred in those instances, when the
communication function of the utterance is to emphasize the contrast between an
enormous effort and zero result. This conclusion is supported by the most frequent
right-hand collocations of the conjunctional expression čo ako (verbs of volition
usilovať sa, snažiť sa, chcieť) and intensifiers of action (Ale čo ako húževnato by ste
okolo neho magnetom čarovali, čo aj tým najsilnejším, aký len existuje, o akom sa
Mesmerovi ani nesnívalo, sval sa nepohne.).
The conjunction akokoľvek is functionally close to the conjunctional expression
čo ako because it highlights the contrast between quite a high but not sufficient
degree of a feature (Akokoľvek budú principiálni, ich jedinou možnosťou je
momentálne politika kompromisu.).
Operators for Extending and Developing an Utterance
195
Utterances with the conjunctional expression ani keď present disappointment
raising from the fact that something (some action or state)does not happen or occur
even under extreme or extraordinary circumstances (Keď ako štyridsaťpäťročný máte
vzrast priemerného dieťa, nič na tom nezmeníte, ani keď budete mať
deväťdesiatpäť.)
Even a higher degree of contrast and disappointed expectations is reached in
utterances with the operators i keby/aj keby presenting the fact that action/state does
not happen/occur even if an unreal circumstance becomes a real one (Nemôže sa
naučiť algebru, i keby mal úžasného učiteľa. – Už nemôže nič povedať, nič
dovysvetľovať, aj keby sa na to odvážila.).
The conjunctional expressions navzdory tomu, že – bez ohľadu na to, že –
nehľadiac na to, že create utterances whose communication intention is to ensure the
addressee aboutthe truth value of the proposition even if this seems to be surprising,
improbable, incompatible with his/her experience or inconsistent with conventions.
Their communication function can be described as hoci + insistent ensuring of the
addressee about the truth value of the proposition.
(14a) Ak poviete v New Yorku meno Andy, nehľadiac na to, že tam môže žiť 300
tisíc mužov s týmto menom, každý dodá Warhol.
(14b) Pri vstupe na štadión, bez ohľadu na to, že v minulosti bol najväčšou
hviezdou domáceho trávnika, zaplatil si vstupné.
(14c) Felipe sa napokon definitívne rozhodol pre Letiziu, navzdory tomu, že nemala
modrú krv.
4.3.3 Finally, the paper focuses on concessive operators in utterances which have the
same formal structure as “real” concessive utterances, but regarding the relation
between their propositions, they do not match a negative implication x, y (thus it is
not a relation between ineffective cause and unrealized consequence). These
utterances are pragmatic and express speaker´s evaluating comment “I am aware of
the fact that what I am going to say will raise doubts”. In this way the speaker
indicates that the proposition which he/she is going to say is not generally
acceptable and he/she counts with a potential accusation of breaking utterance
quality rules or principles of courtesy (Hoci je to paradoxné,... – Akokoľvek
paradoxne to znie,... – Hoci to môže byť trochu pritiahnuté za vlasy,... – Hoci je to
zvláštne,... – Hoci ide o otrepanú frázu,... – Hoci na prvý pohľad tieto dve veci
spolu nesúvisia,... – Akokoľvek sa to zdá absurdné,...). These utterances are close
to the concessive relation because speaker´s doubtful evaluating comment turns out
to be baseless in the further part of text. In my opinion, the category of quality
motivates also utterances in which the concessive operator introduces a comment
related to the speaker´s professional competence (Hoc ekológia nie je mojou
parketou, napadá mi..., – Hoci nie som odborník na trestné právo, chcem uviesť...).
196
Jana Kesselová
5 Conclusion
The aim of this communicative-functional analysis is to depict dynamics within the
micro-system of concessive operators. This dynamics is conditioned by different
communication intentions of language users. Contemporary written Slovak has an
extraordinarily miscellaneous and formally, stylistically, distributionally and
pragmatically differentiated inventory of concessive operators. Dynamics in this
microsystem is demonstrated by strengthening the position of a prototype
representative and identifying the centre with synchronically most frequent
operators on the one hand, and decline (even disappearance) of some synonymous
operators on the other hand. Anothermanifestation of dynamics is a stylistic and
pragmatic specification of operators which is not connected to the operator as a
whole, but to its position in the utterance (initial or non-initial). Concessive
operators are used mainly in belletristic and publicistic styles, less in a scientific
style. The relation of negative implication which is the basis of concessive relation
semantics is used pragmatically in many ways: in utterances with various communication functions such as statement, opinion, justification, evaluation, critique,
defence against critique, ensuring of the addressee, surprise, astonishment,
disappointment and disillusionment.
Author of the paper would like to thank to grant commission VEGA for
supporting the project No. 1/0169/09 which allows realization of analyses presented
in this paper.
References
[1] Dolník, J. – Bajzíková, E. (1998). Textová lingvistika. Bratislava: Stimul.
134 p.
[2] Dolník, J. (1999). Základy lingvistiky. Bratislava: Stimul. 228 p.
[3] Grepl. M. – Karlík, P. (1986). Skladba spisovné češtiny. Praha: Státní
pedagogické nakladatelství. 474 p.
[4] Hirschová, M. (2006). Pragmatika v češtine. Olomouc: Univerzita Palackého.
243 p.
[5] Karlík, P. (1995). Studie o českém souvětí. Brno: Masarykova univerzita.
135 p.
[6] Kolokace. Studie z korpusové lingvistiky. (2006). Eds.: F. Čermák a M. Šulc.
Praha: Nakladatelství Lidové noviny. 452 p.
[7] Korpusová lingvistika. Stav a modelové přístupy. (2006). Eds.: F. Čermák
a R. Blatná. Praha: Nakladatelství Lidové noviny. 358 p.
[8] Müllerová, O. (1979). Komunikativní složky výstavby dialogického textu.
Praha: Univerzita Karlova.. 161 p.
[9] Morfológia slovenského jazyka. (1966). Red. J. Ružička. 1. vyd. Bratislava:
Vydavateľstvo Slovenskej akadémie vied. 895 p.
[10] Oravec, J. – Bajzíková, E. (1982). Súčasný slovenský spisovný jazyk. Syntax.
Bratislava: Slovenské pedagogické nakladateľstvo. p. 171 – 175.
Operators for Extending and Developing an Utterance
197
[11] Pauliny, E. (1981). Slovenská gramatika. Bratislava: Slovenské pedagogické
nakladateľstvo. 288 p.
[12] Slovenský národný korpus – prim-4.0-public-all. Bratislava: Jazykovedný
ústav Ľ. Štúra SAV 2009. Dostupný z WWW:
http://korpus.juls.savba.sk.
[13] Slovenský národný korpus – prim-4.0-vyv. Bratislava: Jazykovedný ústav
Ľ. Štúra SAV 2009. Dostupný z WWW:
http://korpus.juls.savba.sk.
[14] Slovník slovenského jazyka. III. (1963). Vedecký redaktor Š. Peciar. Bratislava: Vydavateľstvo Slovenskej akadémie vied. 911 p.
[15] Slančová, D. (1994). Praktická štylistika. Prešov: Slovacontact. 178 p.
[16] Svozilová, H. (1986). Spojky a další spojovací prostředky. In: Mluvnice
češtiny. 2. Tvarosloví. Praha: Academia. p. 214 – 227.
[17] Šulc, M. (2006). Frekvence javu v korpusu a dva typy jejího referenčního
rámce. In: Korpusová lingvistika: Stav a modelové přístupy. Eds.: F. Čermák,
R. Blatná. Praha: Nakladatelství Lidové noviny. p. 330 – 346.
Changes in Valency Structure of Verbs:
Grammar vs. Lexicon⋆
Václava Kettnerová and Markéta Lopatková
Institute of Formal and Applied Linguistics
Charles University in Prague, Czech Republic
Abstract. In this paper, we deal with changes in valency structure of Czech verbs
from a lexicographic point of view. We focus only on syntactic constructions that
are related in principle to the same (generalized) situation. Changes in valency
structure are understood as different mappings between individual participants of
a generalized situation and valency slots, including their morphemic realization.
We distinguish two types of changes in valency structure, so-called grammatical
diatheses and semantic diatheses. We introduce a basic typology of potential
changes in valency structure and we propose a method of the representation of
these changes in the valency lexicon of Czech verbs VALLEX.
1 Motivation
Syntactic behavior of verbs is determined to a great extent by their lexical semantic
properties. Prototypically, a single valency structure corresponds to a single meaning
of verb. However, in many cases semantically related uses of verbs can be syntactically
structured in different ways. E.g., the pairs of sentences in (1a)-(1b), (1a)-(2a) and (1b)(2b) differ in their syntactic structure despite their obvious semantic similarity:
(1) a. Peter loaded the truck with hay. — b. Peter loaded hay on the truck.
(2) a. The truck was loaded with hay. — b. Hay was loaded on the truck.
Such uses of the verb load cannot be described by a single valency frame; however,
separating four valency frames appears to be redundant with respect to the regularity in
morphemic realizations of valency slots. Let us focus on the pairs of sentences (1a)-(2a)
and (1b)-(2b). In these cases, (i) the information on the possibility of such change in
valency structure of the verb load and (ii) the rule describing such change are sufficient
for lexicographic description. Other changes in valency structure of verbs can be treated
in a similar way under the condition that these changes are so regular that they can be
captured by means of rules.
In this contribution, we deal with changes in valency structure of Czech verbs
from a lexicographic point of view. We introduce and exemplify a basic typology of
potential changes in valency structure of Czech verbs as they have appeared during the
lexicographic processing language data (based on corpus evidence). Finally, we propose
a method of representing these changes in a valency lexicon of Czech verbs.
⋆
The research is carried under the MŠMT ČR project No. MSM0021620838 and partially under
the MŠMT grant No. LC536 and GA UK grant No. 7982/2007.
Changes in Valency Structure of Verbs
199
Basic approaches to changes in valency structure. In Czech linguistics, the study
of syntactic constructions characterized by changes in valency structure of verbs from
the syntactic point of view started in the late sixties, mainly under the influence of
Russian linguistics, esp. [1, 3, 6]. The terms hierachization, diathesis or conversion were
introduced in Czech and Slovak grammars, see esp. [7, 8, 15, 21] and [11]. Roughly
speaking, such terms refer to change in mutual assignment of semantic participants and
(surface) syntactic positions, while the real situation expressed by sentences remains
the same.
In American linguistics, there are three basic approaches to changes in valency structure of verbs, (i) structurally based approaches represented mainly by
transformational-generative grammars, esp. [4, 5], (ii) lexically based approaches focusing on the relation between lexical semantic properties of verbs and their syntactic
behavior, esp. [12], and (iii) constructionally based approaches based on the assumption
that difference in syntactic forms marks the difference in meaning, esp. [2, 10].
Here we focus on the description of changes in valency structure of verbs in the
theoretical framework of the Functional Generative Description (FGD), see esp. [20].
The valency theory of FGD, esp. [16], was applied to a large number of data in building the Prague Dependency Treebank, PDT 2.01 and the valency lexicon of Czech
verbs, VALLEX2 [13]. We attempt to propose an adequate framework for description of
changes in valency structure of verbs which can be applied in lexicographic processing
of language data.
2 Basic typology of changes in valency structure of verbs
In our typology of changes in valency structure of verbs, the concept of situation plays
a key role. The (generalized) situation represents a class of abstract situations characterized by a particular set of semantic participants.3 In the present paper, we focus only
on those syntactic constructions that relate to the same (generalized) situation. Such a
situation is expressed by a single verb lexeme and it is characterized by an identical
set of semantic participants. Changes in valency structure are understood as different
mappings between individual semantic participants of a generalized situation and their
surface syntactic positions, including their morphemic realization. We distinguish two
types of changes in valency structure, so-called grammatical diatheses (g-diatheses) and
semantic diatheses (s-diatheses).
2.1 Grammatical diatheses
G-diatheses represent pairs of related syntactic constructions that prototypically satisfy
the following criteria:
1
2
3
http://ufal.mff.cuni.cz/pdt2.0/
http://ufal.mff.cuni.cz/vallex/2.5/
See also type situation [8, 22] or semantic event. Semantic participants roughly correspond to
semantic roles here.
200
Václava Kettnerová and Markéta Lopatková
I. Verbs in the marked construction are prototypically morphologically marked with
respect to the grammatical category of voice. Their forms typically either consist
of auxiliaries and non-finite form of lexical verbs or they have reflexive forms.
II. The mapping between semantic participants of a generalized situation and valency
slots remains unchanged, their number and type are identical as well. Changes
in valency frames are typically connected with a choice of a particular valency
member for the subject syntactic positions; these changes are limited to morphemic
realizations of individual valency slots.
G-diatheses primarily represent a language means that enables the speaker to choose
a particular semantic participant of a generalized situation for the syntactically prominent position of (surface) subject. In the marked case, the valency member ACT (Actor,
corresponding to the semantic participants of generalized situation such as Agent, Initiator, Causator, Bearer of Action, etc.) is prototypically shifted from the subject syntactic position into a less prominent surface position; eventually, it cannot be expressed
on the surface syntactic level at all (as in deagentive g-diathesis, see e.g. [9]). Another
semantic participant of a generalized situation (typically having the form of accusative)
is shifted into the subject syntactic position, as in (1a)-(2a) repeated below.4 Under
certain conditions, a ‘subject-less’ construction occurs (see example (7b) below).
(1) a. Peter.ACT loaded the truck.PAT with hay.EFF
(2) a. The truck.PAT was loaded with hay.EFF (by Peter.ACT)
G-diatheses can be illustrated by the scheme in Figure 1, the asymmetry concerns
the different mappings between a set of valency members and their surface positions.
Fig. 1. Mapping between semantic participants of a generalized situation and their surface syntactic positions for passive diathesis as a typical g-diathesis (for the verb naložit ‘to load’).
We assume that changes in the valency structure of verbs characteristic of g-diatheses
arise from the special verbal meanings. These verbal meanings are reflected as values
of relevant verbal grammatemes in FGD (grammatemes represent tectogrammatical
correlates of the morphological categories, see [14, 19]).
4
We mark the valency members with labels (so-called functors) ACT, PAT, EFF etc. in accordance with FGD (and with VALLEX in particular).
Changes in Valency Structure of Verbs
201
2.2 Semantic diatheses
S-diatheses are characterized by changes in number and type of valency slots, while the
(generalized) situation still remains unchanged. Furthermore, verbs are not morphologically marked with regard to voice. Contrary to g-diatheses, it is not apparent which of
the related constructions should be understood as unmarked ones and which as marked
ones, see also [8].
Moreover, s-diatheses are typically associated with coherent semantic classes of
verbs, as in sentences (1a)-(1b) (see also, e.g., spray/load verbs in [12]).
(1) a. Peter.ACT-Agent loaded the truck.PAT-Container
with hay.EFF-Filler
b. Peter.ACT-Agent loaded hay.PAT-Filler
on the truck.DIR-Container
In Czech grammars, s-diatheses are described as hierarchizations without marked
voice [8], as objective diatheses [11], or some of them are treated as examples of the
so-called decauzativization [11].
S-diatheses can be illustrated by the scheme in Figure 2, the asymmetry concerns
the different mappings between a set of semantic participants of a generalized situation
and a set of valency members.
Fig. 2. Mapping between semantic participants of a generalized situation and their surface syntactic positions for Container-Filler diathesis (for the verb naložit ‘to load’).
As to the possibility of combining g- and s-diatheses, diatheses of different types are
mutually combinable; i.e., having a marked construction with respect to a g-diathesis, a
particular s-diathesis rule may be subsequently used (if applicable for the given verb),
and conversely, see ex. (1)-(2) in Section 1. However, mutually combining diatheses of
the same type is very restricted.5
Distinguishing between g-diatheses and s-diatheses is motivated by the needs of
lexicographic work. We will see later that in case of g-diatheses, the changes in valency
5
E.g., Když se dostane přidělena pracovna, to se to pracuje. — Eng. If a new study is allocated,
it is easy to work (example from [9]).
202
Václava Kettnerová and Markéta Lopatková
frames are regular enough to be treated within a single verbal lexical unit – general rules
in the grammar component and information on their applicability to individual lexical
units in the data component of the lexicon are sufficient. However, for s-diatheses,
we propose to set separate lexical units interlinked with general rules identifying a
relevant type of s-diathesis. This solution results from the corpus evidence that changes
in valency structure of verbs are diverse even within an individual type of s-diatheses.
3 Representation of G-diatheses
In this section, we introduce a way of capturing g-diatheses in the valency lexicon
VALLEX. In our approach, g-diatheses are described by means of general fine-grained
rules in the grammar component of the valency lexicon. All applicable g-diatheses are
listed for each verbal lexical unit separately in a special attribute in the data component
of the lexicon.
Our method will be demonstrated on the passive diathesis as a prototypical
g-diathesis. Deagentive diathesis, recipient diathesis, resultative diathesis and
mediopassive diathesis, see esp. [19], can be described in the same way. In addition, we
consider also reciprocity as a phenomenon that can be treated in a similar way (within
FGD, reciprocity and the possibility of its representation have been broadly studied by
Panevová, esp. [17]).6
3.1 Passive diathesis
Passive diathesis is a relation between two syntactic constructions in which the marked
one contains the auxiliary verb být ‘to be’ and the past participle of a lexical verb. We
propose the following representation of passive diathesis in the valency lexicon:
(i) In the data component, a single lexical unit is represented by an (unmarked)
valency frame. If a given lexical unit can be subject to passive diathesis, then its
applicability is indicated in the special attribute ‘diathesis-pass’.
(i) In the grammar component, a general rule describing regular changes in a valency
frame for this diathesis is stored.
For example, a lexical unit for the transitive verb postavit ‘to build’ has three valency
slots in its valency frame: obligatory ACT (Actor, in nominative in the unmarked construction), obligatory PAT (Patient, in accusative) and optional ORIG (Origin, expressed
as the prepositional group z ‘from/of’ plus genitive). In the marked construction, ACT
is realized either as instrumental or as prepositional group od ‘by’ plus genitive, and
PAT is expressed as nominative (morphemic realization of ORIG remains unchanged):
(3) a. David.ACTnom postavil kůlnu.PATacc ze dřeva.ORIGz+gen
Eng. David.ACT built a shed.PAT from wood.ORIG
b. Kůlna.PATnom byla postavena ze dřeva.ORIGz+gen (Davidem / od Davida.ACTinstr,od+gen)
Eng. A shed.PAT was built from wood.ORIG (by David.ACT)
6
Causative constructions are another candidates that can be taken into account for this type of
representation.
Changes in Valency Structure of Verbs
203
Passive diathesis for verbs with valency member expressed by accusative. Passive
diathesis concerns verbs with at least two semantic participants of a generalized situation and thus at least two valency slots, prototypically ACT in nominative and PAT in
accusative. Valency frame for the marked member of the diathesis can be described by
the following rule Pass.r1.PAT, see Table 1.
It should be stressed here that all information captured in valency frame remains
unchanged, unless a change is explicitly mentioned by the rule Pass.r1.PAT; i.e., if a
valency frame contains a member or morphemic form that is not cited in the rule, then
it is preserved also in a derived valency frame.
Pass.r1.PAT
Unmarked
verbal grammateme diathesis-pass: 0
valency frame
ACTnom
PATacc
PATvar,inf ,dcc
? EFFjako+acc
Marked
Note
diathesis-pass: pass (1)
ACTinstr,od+gen
(2)
PATnom
(3)
PATexcluded
(4)
? EFFjako+nom
(5)
Table 1. Pass.r1.PAT rule for the passive diathesis.
Commentary on the Pass.r1.PAT rule:
(1) The passive diathesis is represented by the verbal grammateme ‘diathesis-pass’; its value for
the unmarked member of the pair is ‘0’, for the marked member it is ‘pass’.
(2) In the marked construction, ACT is shifted from the prominent subject syntactic position into
the adverbial position. This change is accompanied by the change of morphemic realization of
ACT from nominative into instrumental or into the prepositional case od ‘by’+genitive.
(3) The valency member PAT (expressed by accusative) is selected for the prominent surface
syntactic position of subject for the marked member of the passive diathesis. Its morphemic form
is changed into nominative.
(4) If the PAT valency member may be expressed also by other morphemic forms such as infinitive
(abbr. inf), dependent content clause (dcc) or another preposition or prepositionless case (var)
(mentioned below as ‘unaccusative variants’), all these possible morphemic variants are excluded
in the marked frame. PAT expressed by unaccusative forms is treated with Pass.r2.PAT rule, see
below.
(5) If there is a slot for EFF in the unmarked frame with the form jako ‘as’+accusative, then its
form is changed into jako ‘as’+nominative.
Note on agreement: Verbal categories of person, number and gender agree with ACT in nominative in the unmarked construction, whereas a verb in the marked construction has agreement with
PAT in nominative.
For example, by applying Pass.r1.PAT rule to the unmarked valency frame for
the verb postavit ‘to build’, see ex. (3a)-(3b), we obtain the following valency frame
describing the marked syntactic construction:
ACTnom PATacc ORIGz+gen
⇒ Pass.r1.PAT ACTinstr,od+gen PATnom ORIGz+gen
204
Václava Kettnerová and Markéta Lopatková
The change in the realization of EFF expressed with jako ‘as’+accusative may be
exemplified by the verb hodnotit ‘to assess’. See the unmarked and marked valency
frames and their realizations in sentences (4a)-(4b) (note also the reduction of possible
morphemic forms for PAT in (4b)):
ACTnom PATacc,var,inf ,dcc EFFjako+acc,na+acc
⇒ Pass.r1.PAT ACTinstr,od+gen PATnom EFFjako+nom,na+acc
(4) a. Učitelé.ACTnom hodnotili jeho práci.PATacc
jako nedostatečnou.EFFas+acc
Eng. The teachers.ACT assessed his paper.PAT as poor.EFF
b. Jeho práce.PATnom byla hodnocena učiteli.ACTinstr jako nedostatečná.EFFas+nom
Eng. His paper.PAT was assessed as poor.EFF by his teachers.ACT
For some verbs with at least three valency members, the accusative position may be
labeled with other functors, namely ADDR (for Addressee) or EFF (for Effect),7 see
(5a)-(5b) and (6a)-(6b). The changes in valency structure of these verbs are captured by
analogous rules Pass.r1.ADDR and Pass.r1.EFF.
(5) a. Sekretářka.ACTnom ředitele.ADDRacc upozornila, (že má podepsat
smlouvu).PATdcc
Eng. The secretary.ACT has reminded the director.ADDR (to sign
the contract).PAT
b. Ředitel.ADDRnom byl upozorněn sekretářkou.ACTinstr , (že má podepsat smlouvu).PATdcc
Eng. The director.ADDR has been reminded by his secretary.ACT
(to sign the contract).PAT
(6) a. Zadržený.ACTnom řekl vyšetřovateli.ADDRdat lež.EFFacc
Eng. The detained man.ACT said to the interrogator.ADDR a lie.EFF
b. Vyšetřovateli.ADDRdat byla (zadrženým.ACTinstr ) řečena lež.EFFnom
Eng. A lie.EFF was said to the interrogator.ADDR (by the detained
man.ACT)
Passive diathesis for verbs with valency member expressed by ‘unaccusative’ forms.
Furthermore, passive diathesis can be applied to verbs with valency members realized
by ‘unaccusative’ forms, see ex. (7a)-(7b):
(7) a. Radní.ACTnom o té záležitosti.PATo+loc rozhodli včera.
Eng. The councilors.ACT decided the matter.PAT yesterday.
b. O té záležitosti.PATo+loc bylo (radními.ACTinstr ) rozhodnuto včera.
Eng. The matter.PAT was decided (by councilors.ACT) yesterday.
Changes in valency frame are described by the following rule Pass.r2.PAT, see
Table 2. Again, except for the changes explicitly mentioned in the rule, all other information captured in a valency frame remains unchanged.
7
We leave aside the functors DPHR (for Dependent Part of Phraseme) and CPHR (Part of
Compound Predicate) here.
Changes in Valency Structure of Verbs
Pass.r2.PAT
Unmarked
verbal grammateme diathesis-pass: 0
valency frame
ACTnom
PATvar,inf ,dcc
? PAT|ADDR|EFFacc
205
Marked
Note
diathesis-pass: pass
(1)
ACTinstr,od+gen
(2)
PATvar,inf ,dcc
(3)
? PAT|ADDR|EFFexcluded (4)
Table 2. Pass.r2.PAT rule for the passive diathesis.
Commentary on the Pass.r2.PAT rule:
(1) and (2) See the Commentary on the Pass.r1 rule.
(3) The ‘unaccusative’ morphemic realization of PAT8 remains unchanged. If PAT is realized by
infinitive or dependent content clause, it is shifted into the subject syntactic position. Applying
the given rule to PAT expressed by prepositional case or prepositionless case (with the exception
of accusative), ‘subject-less’ sentence is created.
(4) The possible accusative realization of any valency slot is excluded. If no other morphemic
variant remains, the given valency member cannot be realized in a surface sentence,9 see also ex.
(8c).
Note on agreement: In the marked construction, verbs have incongruent agreement with 3rd sg.
neutr.
Let us exemplify the application of Pass.r2.PAT rule to the valency frame of the verb
rozhodnout ‘to decide’, see also sentences (7a)-(7b):
ACTnom PATo+loc,dcc ⇒ Pass.r2.PAT ACTinstr PATo+loc,dcc
Verbs allowing for two passive constructions. There are verbs allowing for two
passive constructions. First, such verb has a valency member that may be realized
both as accusative and ‘unaccusative’ form (e.g., the verb hodnotit ‘to asses’, see ex.
(4)) – then both types of rules are applicable to this valency member (Pass.r1.PAT
or Pass.r2.PAT for the verb hodnotit ‘to asses’). The second case is represented by
verbs with at least three semantic participants of generalized situations. Such verbs
have at least three valency members (prototypically realized as nominative, accusative
and ‘unaccusative’).10 Again, both types of rules may be used – they are applied to two
different valency members depending on the choice of subject. We exemplify this by the
verb žádat ‘to ask’, see sentence (8a) for the unmarked case, (8b) for the Pass.r1.ADDR
rule and (8c) for the Pass.r2.PAT rule:
ACTnom ADDRacc PATo+acc,inf ,dcc
⇒ Pass.r1.ADDR ACTinstr,od+gen ADDRnom PATo+acc,inf ,dcc
ACTnom ADDRacc PATo+acc,inf ,dcc
⇒ Pass.r2.PAT ACTinstr,od+gen ADDRgeneral PATo+acc,inf ,dcc
8
9
10
The analogous rules are set for ADDR and EFF.
This case results in so called generalized valency member in FGD, see [18].
The verb učit ‘to teach’ with two valency members expressed in accusative represents a rare
exception.
206
Václava Kettnerová and Markéta Lopatková
As the accusative is the only possible realization of ADDR in the unmarked valency
slot (and accusative is excluded in the marked valency frame according to Pass.r2.PAT
rule), the ADDR valency slot cannot be realized in the surface sentence, see ex. (8c).
(8) a. Novináři.ACTnom vládu.ADDRacc žádali, (aby byly zveřejněny výsledky).PATdcc
Eng. The journalists.ACT asked the government.ADDR (to publish
the results).PAT
b. Vláda.ADDRnom byla (novináři.ACTinstr ) žádána, (aby byly zveřejněny výsledky).PATdcc
Eng. The government.ADDR was asked (by the journalists.ACT) (to
publish the results).PAT
c. Novináři.ACTinstr bylo opakovaně žádáno, (aby byly zveřejněny
výsledky).PATdcc (general ADDR)
‘(by) journalists - was - repeatedly - asked - to - publish - results’ Eng. The
publication of the results was repeatedly asked (by the journalists).
4 Representation of S-diatheses
In this section, we focus on s-diatheses and their adequate representation in the valency lexicon VALLEX. To recapitulate, s-diathesis is a relation between two (or more)
syntactic constructions describing a same generalized situation. These constructions
refer to the same (polysemous) verb lexeme, however, the mappings between individual
semantic participants of the generalized situation and valency slots is different. As a
consequence, not only morphemic realization but also number, type and obligatoriness
of valency members may differ. In contrast to g-diatheses, morphological categories of
the given verb typically remain unchanged.
Let us demonstrate our approach on the Container-Filler diathesis as a prototypical
s-diathesis. Other s-diatheses can be captured in the same way (selected examples are
listed below).
4.1 Container-Filler diathesis
Container-Filler diathesis11 can be exemplified by sentences (9a)-(9b) (note that ‘negative’ variant can be also distinguished).
(9) a. Petr.ACTnom -Agent naložil vůz.PATacc -Container
senem.EFFinstr -Filler
Eng. Petr.ACT-Agent loaded the truck.PAT-Container
with hay.EFF-Filler
b. Petr.ACTnom -Agent naložil seno.PATacc -Filler
11
This type of diathesis counts among a group of ‘co-occurrence diathesis’ in [8]; see also
‘spray/load alternation’ in [12]. We adopt a labeling based on semantic participants involved
in the diatheses as we consider it more transparent.
Changes in Valency Structure of Verbs
207
na vůz.DIR-Container
Eng. Petr.ACT-Agent loaded hay.PAT-Filler
on the truck.DIR-Container
These two sentences describe in principle the same generalized situation with three
semantic participants – Agent (who causes the action described by the given verb),
Filler (substance or entity whose location is changed) and Container (location where
Filler is moved). Despite the single set of semantic participants of the generalized
situation, this situation can be structured in a different way. While Agent is realized
as ACT in both cases, there are two possibilities for Filler and Container: (i) either
Container is mapped onto PAT (in accusative) and Filler is mapped onto EFF valency
slot (in instrumental), as in (9a); (ii) or Filler occupies the PAT slot (in accusative) and
Container is structured as Directional modification DIR, as in (9b) (see also Figure 2 in
Section 2.2).
The most studied semantic property of this diathesis deals with a partitive / holistic effect. The semantic participant of the generalized situation realized as PAT in
accusative typically receives holistic interpretation; i.e., in Container-Filler diathesis
either Container (9a) or Filler (9b) is understood as completely affected by the action
expressed by the verb naložit ‘to load’.
Contrary to g-diatheses, the changes in valency frames accompanying s-diatheses
are not regular enough: individual verbs exhibit many irregularities in their valency
characteristics even within a single type of s-diathesis (see below for the examples).
For the purpose of the valency lexicon VALLEX, we propose the following representation of s-diatheses:
(i) In the data component, we establish a set of two lexical units within one lexeme
– each member of s-diathesis is represented by a separate lexical unit with its
own valency frame. These lexical units are interlinked via the type of s-diathesis
(captured in a special attribute ‘s-diathesis’).
(ii) In the grammar component, a general rule describing possible mappings between
semantic participants of a generalized situation and individual valency slots is provided, see Table 3.
Container-Filler
Filler ∼ PAT
Agent Filler Container examples
ACT
PAT DIR
naložit seno na vůz
doplnit cukr do cukřenky
nasypat mouku do pytle
(na)točit vodu (do kýble)
Container ∼ PAT ACT
EFF PAT
naložit vůz senem
doplnit cukřenku cukrem/o cukr
ACT
—
PAT
nasypat pytel *moukou
(na)točit kýbl *vodou
Table 3. General rule for the Container-Filler diathesis (see the translations below).
208
Václava Kettnerová and Markéta Lopatková
The dissimilarities in the Container-Filler diathesis concern number, type, and morphemic realization of complements as well:
– Whereas the set of semantic participants of the generalized situation is the same
(Agent, Filler, Container) and prototypically all of them can be realized as valency
members, this does not hold for some verbs (e.g., nasypat mouku do pytle ‘to put
flour into the sack’ but nasypat pytel *moukou ‘to put the sack *with flour’).
– Whereas directional valency member that realizes Container participant is prototypically obligatory (e.g., doplnit cukr do cukřenky ‘to add sugar to the sugar bowl’),
there are verbs with only typical directional valency member (e.g., točit vodu (do
kýble) ‘to draw water (to the bucket)’).
– Morphemic realizations of a particular valency member may differ with individual
verbs (e.g., doplnit cukřenku cukrem / o cukr ‘to replenish the sugar bowl with
sugar’).
4.2 Examples of other S-diatheses
While g-diatheses are intensively studied in Czech linguistics, there is only a limited
number of studies of phenomena referred here to as s-diatheses, see esp. [8]. Let us
exemplify here at least several frequent s-diatheses in Czech which can be captured in
the valency lexicon in a similar way as the Container-Filler diathesis:
Surface-Cover diathesis (positive or negative)
Jana si očistila bláto.PAT-Cover z bot.DIR-Surface
Eng. Jane cleaned the mud.PAT-Cover off her shoes.DIR-Surface
— Jana si očistila boty.PAT-Surface od bláta.ORIG-Cover
Eng. Jane cleaned her shoes.PAT-Surface of the mud.ORIG-Cover
Material-Product diathesis (positive or negative)
Kadeřník jí učesal vlasy.PAT-Material do drdolu.EFF-Product
Eng. Hairdresser arranged her hair.PAT-Material into a bun.EFF-Product
— Kadeřník jí učesal z vlasů.ORIG-Material drdol.PAT-Product
Eng. Hairdresser arranged a bun.PAT-Product from her hair.ORIG-Material
Source-Substance diathesis
Slunce.ACT-Source vyzařuje teplo.PAT-Substance
Eng. The sun.ACT-Source radiates heat.PAT-Substance
— Teplo.ACT-Substance vyzařuje ze slunce.DIR-Source
Eng. Heat.ACT-Substance radiates from the sun.DIR-Source
Object-Direction diathesis (‘from where’, ‘through’ or ‘to where’)
Marta vylezla kopec.PAT-Object
Eng. Martha climbed the mountain.PAT-Object
— Marta vylezla na kopec.DIR-Direction
Eng. Martha climbed up the mountain.DIR-Direction
Direction-Location diathesis
Matka umístila dítě do jeslí.DIR-Direction
Eng. Mother put her child into a nursery school.DIR-Direction
— Matka umístila dítě v jeslích.LOC-Location
Eng. Mother put her child into a nursery school.LOC-Location
Changes in Valency Structure of Verbs
209
Agent-Location diathesis
Včely.ACT-Agent se rojí na zahradě.LOC-Location
Eng. Bees.ACT-Agent are swarming in the garden.LOC-Location
— Zahrada.ACT-Location se rojí včelami.MEANS-Agent
Eng. The garden.ACT-Location is swarming with bees.MEANS-Agent
Conclusion
For lexicographic description of verbal valency, it is necessary to specify (i) valency
frame of each lexical unit, (ii) information on the applicability of a particular set of rules
describing the possible diatheses, and (iii) precise formulations of rules. Information
(i) and (ii) are stored in the data component whereas (iii) is stored in the grammar
component of the valency lexicon.
We distinguish two types of changes in valency structure, which are referred to as
g-diatheses and s-diatheses. G-diatheses are prototypically characterized by morphologically marked form of verb in the marked construction, while the mapping between
semantic participants of a generalized situation and valency slots remains unchanged,
their number and type are identical (the changes in valency frames are limited to morphemic realizations of individual valency slots). On the other hand, s-diatheses are
characterized by changes in number and types of valency slots. They are typically
limited to verbs of certain semantic classes.
Distinguishing between g-diatheses and s-diatheses in the valency lexicon VALLEX
is motivated by the needs of lexicographic work. In case of g-diatheses, the changes in
valency frames are regular enough to be treated in the form of general rules (in the
grammar component) and as a single verbal lexical unit (for both syntactic constructions) marked with the possibility of a particular type of diathesis. For s-diatheses,
separate lexical units are established and interlinked with general rules identifying a
relevant type of s-diathesis. This solution reflects the corpus evidence that changes in
valency structure of verbs are diverse even within an individual type of s-diathesis.
References
[1] Apresjan, J. D. (1974). Leksicheskaja semantika. Sinonimicheskie sredstva jazyka.
Nauka, Moskva.
[2] Borer, H. (2005). The Normal Course of Events. Oxford University Press, Oxford.
[3] Cholodovič, A. A. (1970). Zalog. Kategoria zaloga. In Materialy konferencii,
pages 2–26, Leningrad.
[4] Chomsky, N. A. (1957). Syntactic Structures. Mouton, The Hague.
[5] Chomsky, N. A. (1965). Aspects of the Theory of Syntax. MIT Press, Cambridge.
[6] Chrakovskij, V. S., editor (1977). Problemy lingvisticheskoj tipologii i struktury
jazyka. Nauka, Leningrad.
[7] Daneš, F. (1968). Some Thoughts on the Semantic Structure of the Sentence.
Lingua, 21:55–69.
[8] Daneš, F. (1985). Věta a text: studie ze syntaxe současné češtiny. Academia,
Praha.
210
Václava Kettnerová and Markéta Lopatková
[9] Daneš, F., Grepl, M., and Hlavsa, Z., editors (1987). Mluvnice češtiny 3. Academia, Praha.
[10] Goldberg, A. E. (1995). Constructions. A Construction Grammar Approach to
Argument Structure. University of Chicago Press, Chicago.
[11] Grepl, M. and Karlík, P. (1998). Skladba češtiny. Votobia, Olomouc.
[12] Levin, B. C. (1993). English Verb Classes and Alternations: A Preliminary
Investigation. The University of Chicago Press, Chicago and London.
[13] Lopatková, M., Žabokrtský, Z., and Kettnerová, V. (2008). Valenční slovník českých sloves. Nakladatelství Karolinum, Praha.
[14] Mikulová, M. et al. (2006). Annotation on the Tectogrammatical Level in the
Prague Dependency Treebank. Annotation Manual. Technical Report TR-200630, ÚFAL MFF UK, Prague.
[15] Ondrejovič, S. (1989). Medzi slovesom a vetou. Jazykovedné štúdie. Veda, Bratislava.
[16] Panevová, J. (1980). Formy a funkce ve stavbě české věty. Academia, Praha.
[17] Panevová, J. (2007). Znovu o reciprocitě. Slovo a slovesnost, 68:91–100.
[18] Panevová, J. and Řezníčková, V. (2001). K možnému pojetí všeobecnosti aktantu.
In Hladká, Z. and Karlík, P., editors, Čeština – univerzália a specifika 3, pages
139–146. Masarykova Universita, Brno.
[19] Panevová, J. et. al (manuscript). Syntax současné češtiny (na základě anotovaného
korpusu). Nakladatelství Karolinum, Praha.
[20] Sgall, P., Hajičová, E., and Panevová, J. (1986). The Meaning of the Sentence in
Its Semantic and Pragmatic Aspects. Reidel, Dordrecht.
[21] Štícha, F. (1984). Utváření a hierarchizace struktury větného znaku. Univerzita
Karlova, Praha.
[22] Uspenskij, V. A. (1977). K ponjatiju diatezy. In Chrakovskij, V. S., editor, Problemy lingvisticheskoj tipologii i struktury jazyka, pages 65–84. Leningrad.
Corpus-Based Analysis of Lexico-Grammatical Patterns
(on the Corpus of Letters of N. V. Gogol)*
Maria Khokhlova1,2, and Victor Zakharov1,2
2
1
Saint-Petersburg State University
Institute of Linguistic Studies, Russian Academy of Sciences, Russia
Abstract. The paper is concerned with collocation extraction by means of
statistical methods combined with certain syntactic models. Among collocations there is a special type of them which is usually called colligations. It
could be defined as lexico-grammatical patterns, or collocations with regard
to syntactical formulae. There’s a system known as Sketch Engine representing itself a corpus tool which takes as input a corpus of any language and
corresponding grammar patterns. The system gives information about a
word’s collocability on concrete dependency models, and generates lists of
the most frequent phrases for a given word based on appropriate models. The
present paper deals with collocation extraction on the corpus letters of
N.V. Gogol by means of the mentioned system, discusses the results
obtained. The results show that word sketches and information about collocation behaviour could facilitate lexicographic work with the Russian language.
1 Introduction
Nowadays in modern linguistics corpora have turned to be vital tools for linguistic
studies and solution for applied tasks. The application of corpora methods to the
analysis of lexical collocability enables to write grammars and compile dictionaries
of a new type, dictionaries of collocations, idioms etc. With arrival of text corpora
and corpus linguistics lexicographers and other linguists have gained an opportunity
to look at big collections of word usage. Corpora do not only help to study lexical
units in context but also to get data on word frequency, frequency of lexemes, grammatical categories, their collocability etc.
Although the above mentioned corpora opportunities are very useful, there is a
need of another kind of software for further improvement of linguistic research as it
is impossible to process huge amount of linguistic data manually. It can be described
as an additional system between a corpus and its users (linguists) which can process
significant language data.
The objective of such a system is to provide lexicographers with sufficient lexical material and tools for getting information about a word’s collocability and to
generate lists of the most frequent phrases for a given word, and then to classify
them for appropriate syntactic models.
*
The project is supported by Visegrad Scholarship “Syntactic and statistical models of
phrases in Russian” (Contract Number #50810256).
212
Maria Khokhlova and Victor Zakharov
2 Sketch engine
Such a system known as Sketch Engine was developed for a number of languages by
British and Czech scholars (see Sketch Engine project1). The Sketch Engine combines approaches of both traditional linguistics (syntax) and statistics. It is widely
used by scholars in grammar writing and dictionary compiling (Oxford University
Press, Cambridge University Press, Collins, Macmillan etc.).
The Sketch Engine system is developed for a number of European languages
(English, Irish, Spanish, Italian, German, Portuguese, Slovene, French, Czech) and
for the Chinese and Japanese languages. This is a corpus tool which takes as input a
corpus of any language and corresponding grammar patterns and which generates
word sketches for words of that language. Word sketches are one-page automatic,
corpus-based summaries of a word’s grammatical and collocational behaviour [1].
One can understand word sketches as typical phrases determined on the one hand by
syntax that restricts words’ collocability in a given language and on the other hand
by probability closely related to word usage.
3 Word sketch grammar (rules for the Russian language)
The paper presents results of writing rules for the Russian language for the system
Sketch Engine and testing them on the corpus of letters of N. V. Gogol (a famous
Russian writer of the XIXth century).
During our work we have written grammatical rules that take into account syntactic constructions of the Russian language based on the morphologically tagged
corpus in terms of grammar of Sketch Engine. On the basis of these rules and statistical measures the programme generates tables with word sketches for a key word.
While writing rules we used regular expressions and query language IMS Corpus
Workbench. The system searches for tags which correspond to word forms. For
example, tag Ncfpnn means common noun (Nc) female gender (f) plural (p) noun
case (n).
Below there is an example of grammatical rules for the phrases «adjective+noun»:
*DUAL
=a_modifier/modifies
2:"A....n." (([word=","]|[word="и"]|[word="или"]) [tag="A....n."]){0,3} 1:"N...n."
2:"A....g." (([word=","]|[word="и"]|[word="или"]) [tag="A....g."]){0,3} 1:"N...g."
2:"A....d." (([word=","]|[word="и"]|[word="или"]) [tag="A....d."]){0,3} 1:"N...d."
2:"A....a." (([word=","]|[word="и"]|[word="или"]) [tag="A....a."]){0,3} 1:"N...a."
2:"A....i." (([word=","]|[word="и"]|[word="или"]) [tag="A....i."]){0,3} 1:"N...i."
2:"A....l." (([word=","]|[word="и"]|[word="или"]) [tag="A....j."]){0,3} 1:"N...l."
1
http://www.sketchengine.co.uk
Corpus-Based Analysis of Lexico-Grammatical Patterns
213
Above mentioned rules take into account all such phrases, e.g. nouns and adjectives in the same case with conjunctions «и» (“and”), «или» (“or”), comma or
adjectives between them within the distance of 3 words. The numeral 1 stands for
the key word (for instance, 1:"N...n.") and the numeral 2 indicates a collocate (for
instance, 2:"A....n.").
Originally these rules were written on the basis of existing rules for English and
Czech [2].
Then we have written the second variant of word sketches rules within the
approach of Vladimir Benko for the Slovak National Corpus [3].
Its distinctive feature is that these rules describe all phrases found in a corpus.
For example, “verb + any word” (see below):
=Vb X/X Vb
2:[tag="V.*"] 1:[tag!="SENT"]
1:[tag!="SENT"] 2:[tag="V.*"]
The second line means that there will be found all phrases for any word (if it
isn’t a punctuation mark that has its own tag in the corpus) with a verb. The rule in
the third line describes the same phrases but a verb is to the right of the key word.
It should be remarked that this approach has its advantage as word sketches are
generated for any word (because very often morphological ambiguity or mistakes of
automatic tagging prevent from giving objective results when a given word is specified).
As statistical tools in the system there are used the following measures of association: log-likelihood, MI, t-score, MI3, widely used in linguistics for the
computation of the strength of syntagmatic relations.
4 Collocation extraction on the corpus of letters of N. V. Gogol
We have used a collection of letters of N.V. Gogol as test material (about 500 thousand tokens) [4]. Texts were morphologically tagged by the programme TreeTagger
[5]. During this work we have tested the rules and got information about collocability on the corpus data.
Below one can see examples of phrases with the key word «маминька» (“mum”).
214
Maria Khokhlova and Victor Zakharov
Fig. 1. Example of phrases with the key word «маминька» (“mum”)
The first column represents left collocates of the key word «маминька»
(“mum”).
The adjective «Дражайшая» (“dear”, archaic) (the 5th line of a modifier pattern)
was selected as collocate due to the mistake of lemmatization (a form in female gender was ascribed to the word, moreover, as a proper name, this is supported by its
uppercase spelling). The same mistake one can find with the word «Папинька»
(“dad”, archaic).
The second column represents frequency of the given collocate. The third column shows the value of statistical measure according to which the collocate was
selected, it is connected by hyperlink with concordance.
Corpus-Based Analysis of Lexico-Grammatical Patterns
215
It should be stressed that parts of one phrase can be delimited by other words.
For example, for the noun «маминька» there is a verb «обещаться» (“to promise”),
that is located at the distance of two words to the right: Также ежели б еще
прислали чего-нибудь из съестных припасов, как маминька еще тогда
обещались прислать сушеных вишен без косточек [6] (Also if you send some
eatables as in due time mum promised to send some dried seedless cherries). Moreover this example is interesting because in Russian the noun and verb are not
coordinated in number.
Below one can see a part of the table for the verb «иметь» (“to have”).
Fig. 2. Example of phrases with the key word «иметь» (“to have”)
The name of each column reflects syntactical relations (has_obj4 – has an object
in accusative case, has_obj2 – has an object in genitive case, has_subj – has a subject, post_prep – preposition in postposition). As we can see from the first column
the most frequent phrases are «иметь честь» (“to have the honour”), «иметь
обыкновение» (to be in the habit) and «иметь право» (“to have right”): Цалуя без-
216
Maria Khokhlova and Victor Zakharov
ценые ручки ваши имею честь быть, с сыновным моим к вам высокопочитанием ваш послушный сын Николай Гоголь Яновский [7] (Kissing your beloved
hands I have the honour to be with my filial respect to you your obedient son Nikolay Gogol Yanovsky); Я имею право сказать это, как человек, проведший в
борьбе с собой многие годы жизни и лишеньями добившийся до этого права [8]
(I have right to say it as the man who has spent many years wrestling with himself
and owing to hardships has gained this right).
5 Conclusion
There is a question of corpus volume. For example, we know that different association measures extract different collocations but here one can't see differences
between results obtained by a number of statistical measures, it means that collocates will be quite the same. This problem arises from low frequencies of words and
phrases.
If we compare data obtained on the corpus of letters of N.V. Gogol to data
obtained corpora of literary texts, we can often see considerable differences in collocability that reflect author's characteristic word usage. Thus it can be said that
methods described can be effectively used for studying the authors’ language and
writing authors’ dictionaries, for revealing collocability of words in different styles
or within the given time period.
References
[1] Kilgariff, A., Rychlý, P., Smrž, P., Tugwell, D. (2004). The Sketch Engine. In:
Proceedings of EURALEX-2004, 105-116.
[2] Rychlý, P., Smrž, P. Manatee, Bonito and Word Sketches for Czech. (2004). In
Trudy mezhdunarodnoy konferentsii “Korpusnaja lingvistika-2004”: Sbornik
dokladov. St.-Petersburg, 324-334.
[3] Benko, V. Word Sketches for the Slovak National Corpus. In: Mondilex-2009.
Metalanguage and encoding scheme design for digital lexicography (Innovative Solutions for Lexical Entry Design in Slavic Lexicography) Open
Workshop (Bratislava, Slovakia, 15th – 16th April 2009)
[4] Gogol, N.V. Polnoye sobraniye sochineniy: [V 14 t.]. (1937–1952). Moscow –
St.-Petersburg, 1937–1952. T. X-XIV.
[5] http://www.ims.uni-stuttgart.de/projekte/corplex/TreeTagger/
[6] Gogol, N.V. Polnoye sobraniye sochineniy: [V 14 t.]. (1937–1952). Moscow –
St.-Petersburg, 1937–1952. T. X. Pis’ma, 1820–1835, 41.
[7] Gogol, N.V. Polnoye sobraniye sochineniy: [V 14 t.]. (1937–1952). Moscow –
St.-Petersburg, 1937–1952. T. X. Pis’ma, 1820–1835, 31.
[8] Gogol, N.V. Polnoye sobraniye sochineniy: [V 14 t.]. (1937–1952). Moscow –
St.-Petersburg, 1937–1952. T. XII. Pis’ma, 1842–1845, 250.
‘New/novelty’ Concept Set Dynamics as a Marker of
Lexical and Grammatical Paradigm Evolution for
Psychology Sublanguage
Oksana S. Kоzаk
Kyiv National Linguistic University, Ukraine
Abstract. The paper covers on the theoretical foundations and research data
analysis for the specific features of the NEW/NOVELTY concept set dynamics
in the psychology sub-language texts. The analysis has been carried out through
pointing out and statistically studying the set units on the basis of scientific
papers which were published in the 19th and 21st centuries.
Keywords: lexical and grammatical paradigms, a concept set, an invariant, the
centre (nucleus), the periphery, a concept diagram.
The statement that language is experiencing continual changes, i.e. that it is in the
state of dynamics has become a postulate. It is the process of communication that
spots the language dynamics [1, 24]. The written record of communication traces the
state of the language system at a certain stage, the stage being performed in definite
speech situations. This enables linguists to carry out the comparative analysis of the
functional styles on the basis of experimental material where language units get their
embodiment in speech, thus providing the chance of taking extra-linguistic factors
into consideration. Articles in psychology published in scientific proceedings within
two time periods – the second half of the 19th and the beginning of the 21st centuries –
served the material for this research. As both the language structure and its functioning are subordinated to statistical rules, they are to be studied with the help of
statistical methods [2, 7]. The dynamics of functional semantic and stylistic categories
of a certain sub-language can be traced through the changes in lexical and grammatical paradigms. In our opinion, these changes can lead to the changes in the structure
of concepts. The above mentioned dynamics is characterized by the system of markers which in our research are represented by the most frequent concept sets.
A concept set stands for the complex of lexical and grammatical units that serve as
the embodiment for one and the same concept in the texture of text.
The study of concept set dynamics is topical nowadays as it provides the
researcher with the data on the main tendencies and trends in the development of lexical and grammatical paradigm of a certain style sub-language, it also widens the
research horizons for compiling, modifying and updating term vocabularies. This
paper aims at presenting the results of the study on identifying the logic of the
NEW/NOVELTY concept set dynamics via figuring out the changes in the set structure in the texts of the two periods. The concept set that has been defined and studied
within this research, fits into the plane of lexical and grammatical abstraction and cor-
218
Oksana S. Kоzаk
relates with the concept field through the paradigm of lexical and grammatical units.
In their form the elements of the concept set are common root lexemes which stand
for one and the same concept and represent the scope of declension range for the
notional parts of speech (noun, adjective, pronoun, verb and its non-finite forms) [3;
4].
The task of this research has been to define the most frequent units of the
NEW/NOVELTY set, to identify absolute and relative set units frequency for each of
the two periods under research, as well to carry out the analysis of set frequencies
according to their dispersion, range, statistical homogeneity, number of set units.
We assume that the set of the most frequent units of the concept set correlates with
the specific features of the functional semantics and style categories (FSSC) [5, 139]
and their main traits for a certain period.
We state that the dynamics of lexical and grammatical paradigms of the psychology sub-language is the process of its functional and stylistic development,
improvement, thus it results in broadening and specifying the properties of the units
of the above mentioned categories for the best service of the communication aims in
scientific sphere for a certain period.
We have analyzed the texts of Psychology scientific papers with the overall word
usage number of 70 400 (35000 and 35400 word usage cases for the first and the
second period respectively). Thematically related texts for the two periods have been
analyzed. The analysis resulted in defining the most frequent NEW/NOVELTY
concept units for both of the compared text corpora. The total number of units indicates the theme breadth of the articles under study. For Period 1 (the second half of the
19th century) overall units frequency per corpus is 17, for Period 2 (the beginning of
the 21st century) it is 38.
The absolute frequency of concept units has been calculated followed by the analysis of their relative frequencies for the two periods according to:
а) frequency dispersion (the number of texts for each of the periods under research in
which the concept units are represented);
b) frequency range (diapason):
• low (0.01 – 0.09%);
• lower than the medium (0.10 – 0.19%);
• medium (0.20 – 0.29%);
• higher than the medium (0.30 – 0.39%);
• high (0.40% and more);
c) frequency homogeneity (statistically homogeneous/ non-homogeneous frequencies);
d) the quantity of the concept set units for a certain period.
‘New/novelty’ Concept Set Dynamics
219
Here below is the table of comparison for the units frequency of the NEW/NOVELTY concept set for the two periods:
№
1
2
3
4
5
6
7
8
9
10
11
12
PS 1
p = 0.05%
Unit
neonew
newly
news
newsy
renew
renewable
renewal
renewed
novel
novelty
novelties
∑
Frequency
0
13
0
0
0
0
0
0
3
0
1
0
n=17
№
1
2
3
4
5
6
7
8
9
10
11
12
PS 2
p = 0.1%
Unit
neonew
newly
news
newsy
renew
renewable
renewal
renewed
novel
novelty
novelties
∑
Frequency
4
22
0
9
0
0
0
0
0
1
1
1
n=38
Table 1. Correlation between the units quantity for the NEW/NOVELTY
Concept Set (Periods 1, 2)
where PS 1 and PS 2 stand for the first and the second periods of papers on Psychology, m and n – relative and absolute set frequencies respectively, the unit in bold
stands for the core of a concept set of a certain period [6, 107].
Frequency dispersion analysis has provided the possibility to find out that the dispersion of the concept set units is the same. Should the set be represented in the texts
of one of the two periods only, it could be the indicator of increase/decrease in evincing interest for a certain psychological issue.
Frequency range (diapason) analysis has enabled not just to trace the vector of the
NEW/NOVELTY, but has also prompted the idea of building concept diagrams for
both separate articles [5] and for the psychology sub-language in general for each of
the two periods under analysis in accordance with the categories and the core/periphery frequencies.
Frequency homogeneity analysis has made it possible to estimate the statistical
frequency homogeneity/absence of homogeneity for the concept set units of each of
the two periods.
The quantity of the concept set units analysis has enabled to trace the dynamic
changes of lexical and grammatical paradigms within the concept set (the set centre
shift from one unit onto another; increase/decrease in the relevant frequency of the
concept set; the change in lexical and grammatical unit status.
According to the frequency dispersion the NEW/NOVELTY conceptual set is a
prevailing one with a non-significant difference in frequency. According to the fre-
220
Oksana S. Kоzаk
quency range (diapason) the set belongs to different groups for the two different periods: low frequency (0.01 – 0.09%) group (for the first period); lower than the medium
frequency (0.10 – 0.19%) group. According to the frequency homogeneity the
NEW/NOVELTY is a set with statistically non-homogeneous frequencies. There is
also a difference in the qualitative set composition, as well as in its lexical and grammatical status. Although the data from Table 1 do not show the significant shifting of
the concept core (new 13 → new 22), just show its slight movement in the direction of
news 0 → news 9, still the increase in the core units frequency accompanied by an
inconsiderable change in the quantity of units variants in the set periphery for both of
the corpora (novelty 1 → novel/novelty/novelties 3), inspires us to make a conclusion
about the increase in textuality loading on the NEW/NOVELTY concept in the texts
of the Period 2. Moreover, a significant change in the qualitative set composition in
the set periphery has been spotted (neo- 0 → neo- 4; renewed 3 → renewed 0 ). The
results of the analysis are given in Table 2 below:
Period
Dispersion
Period 1
(the second
half of the 19th
century)
Period 2
(the beginning
of the 21st century)
100%
100%
Range
(Diapason)
Low
(0.01 – 0.09%)
Homogeneity
-
Quantity
of Units
3
Lower than the
medium
(0.10 – 0.19%)
Р%
0.03-0.06
-
6
0.05-0.16
Table 2. Unit Frequency for the NEW/NOVELTY Concept Set (Periods 1, 2)
The analysis has enabled to arrive at the conclusion that the concept under
research has a significantly different weight in the texts of the two periods, as well as
to construct a concept diagram for the NEW/NOVELTY set as based on the data
received.
‘New/novelty’ Concept Set Dynamics
221
Chart 1. NEW/NOVELTY Concept Diagram (Periods 1, 2)
There is no doubt that the results of the research prove the need for a further
research at the level of each of the of the functional semantics and style categories of
the psychology sub-language [5], which will make it possible to build the above specified diagrams.
References
[1] Левицький, А. Э. (2002). Особенности динамики лексико-грамматической
системы современного английского языка // Мова і культура (Науковий
щорічний журнал). – К.: Видавничий Дім Дмитра Бураго,. – Випуск 5. – Т
ІІІ. Ч.: Національні мови і культури в їх специфіці і взаємодії. – С.23–28.
[2] Перебийніс В. І. (2001). Статистичні методи для лінгвістів: Навчальний
посібник. – Вінниця: Нова книга. – 168с.
[3] (2003). Longman Dictionary of Contemporary English. – Harlow, England:
Pearson Education Limited. – 1950 p.
[4] Hornby A. S. (2006). Oxford Advanced Learner’s Dictionary. – Oxford: Oxford
University Press.– 1715 p.
[5] Козак О. С. (2006). Мовленнєва системність та стилістичні категорії підмов
філософії та психології // Проблеми семантики, прагматики та когнітивної
лінгвістики: Зб. наук. пр. Вип. 10 / Київ. нац. ун-т ім. Т. Шевченка; Відп.
ред. Н. М. Корбозерова. – К.– С. 139–143.
[6] (2007). Computer Treatment of Slavic and East European Languages // Fourth
International Seminar: Bratislava, Slovakia, 25 – 27 October 2007. – Tribun:
Bratislava. – P. 104 – 108.
Methodological Foundations for Contrastive
Model of Verb Valence
Ružena Kozmová
University of SS. Cyril and Methodius in Trnava, Slovakia
Abstract. Subject of this contribution is to present contrast model of the verb
valence and methodology sources, which are used in the process of model
building.
In the comparison of the verb valence in Slovak and German, we can see
general differences (theoretic basis of the valence, theory of the verb intention,
terminology, structure of the sentence) and specific differences (minor differences in the valence, some differences in the actionality and major differences
in the variations of verb meaning).
In the summary we present sentence models of verb sagen, to which corresponds
Slovak equivalents of perfective verb povedať and imperfective verb hovoriť.
Prologue
Subject of this contribution is presentation of contrastive model of verb valence. We
want to concentrate mainly on the methodological foundations that apply in the model
construction. In the confrontation of verb valence in Slovak language and German
language appears general differences (theoretical basis of the valence, theory of the
verb intention, characteristic of verb and verbal situation, resp. predicates, terminology, model structure) and specific (minor differences in the valence in general,
absence of the aspect category in German and major differences in so called semantic
verb variations).
Attention also deserves clarification of contrastive model of verb valence concept,
which we understand in the sense, that in the model construction of given verb, both
languages are target as well as source. Specifically it means description of sentence
model of verbal lexeme in German language with presentation of its equivalents in
Slovak language. And vice versa model of verbal lexeme in Slovak language and its
equivalents in German language. This procedure is actually contrastive – confrontational and its result is always two models of verbal lexeme. Construction of
contrastive-confrontational model has from the mentioned reason specific character;
therefore it will differ from both German as well as Slovak model of verb valence
used in the existing valence dictionaries1.
1
We have in mind valence models of German verbs beginning with the oldest as is G. Helbig/
Schenkel, W. (1983): Wörterbuch zur Valenz und Distribution deutscher Verben. Tübingen: Max
Niemeyer and ending with the most recent: H. Schumacher/ J. Kubczak/ R. Schmidt, V. de Ruiter
(2004): VALBU – Valenzwörterbuch deutscher Verben. (Studien zur deutschen Sprache 31).
Tübingen: Narr. For Slovak language I quote so far the single valencedictionary: J. Nižníková/
Methodological Foundations for Contrastive Model of Verb Valence
223
1 To theory of valence and intention
Concerning the limitation of space we have to be satisfied with only the outline of the
theory of valence development in the German linguistics, that has been determined
not only by European2, but also by American linguistics. Our goal is to demonstrate,
that two on the outside diverse theories: theory of valence and theory of verb intention
stands in the complementary relation and can create optimal theoretical base for the
characteristics of sentence formula, or sentence model of verb.
Theory of valence, established by Tesnière3, found great response in German linguistics and relatively soon in the frame of its development two streams were
detached according to dominance of linguistics levels. This way, in the beginning of
seventies and eighties a group of linguist was singled out, which was researching
valence on the morphosyntactic base4, mainly under the leadership of G. Helbig (G.
Helbig-M.D. Stepanova, G. HelbigW. / Schenkel etc.) and so called semantic group,
that was represented mainly by K. Welke, K. Bondzio, K.-H. Sommerfeldt. G. Helbig
later5reassessed theoretical foundations of valence in previous period and outlined six
stages complex model of verb valence – model emerging from syntax – because after
all, it is the supplementing of sentence – however, its basis is semantic layer and by
way of semantic markers it is possible to specify more accurate sentence model and
its participants. Sentence model is elementary problem in the composition of sentence
also according to Ružička6 (1968), because sentence and sentence models are important instruments of each statement. It is therefore desirable to look for these criteria in
„features, thus semantic markers of verb as word class.“ For Ružička the most important characteristics of verb are valence and intention of verbal situation, which was
elaborated by Eugen Pauliny7. Such an understanding is in Slavonic8 typical for many
2
3
4
5
6
7
8
Sokolová, M. (1998): Valenčný slovník slovenských slovies. FF Prešov.
Let’s recall from older period J. A. Lehmann and his work„Allgemeiner Mechanismus des
Periodenbaues“ z r. 1833. According to Lehmann, verbum finitum is important from formal
and material aspect, however, subject is from formal aspect irrelevant, but from material
aspect relevant. Model construction works following the principle of dependency: „Subjekt
sei „Nichtnotwendiger oder Nebenbestandteil des ersten Ranges“, other sentence membersare
members of second level, hence „Dependenzen“.
Concept of valence was originally adopted from chemistry and it was taken into linguistics
independently by S. D. Kacneľson (1948) and de Groot (1949). Foundation of the theory was
formulated by L. Tesnière (1959): „Éléments de syntaxe strusturale“.
Compare works of G. Helbiga z r. 1983, 1987.
G. Helbig (1992): Probleme der Valenz- und Kasustheorie. Tübingen: Max Niemeyer.
J. Ružička (1968): In : Jazykovedný časopis.
E. Pauliny (1943): Štruktúra slovenského slovesa. Bratislava: VEDA.
Ľ. Reháček (1966): Sémantika a syntax infinitivu v současném polském spisovném jazyce.
Concept of intention is understood by Reháček as function of lexical semantics of expression
on the contrary to valence, which for him represents only grammatical ability to bind other
grammatical forms.
224
Ružena Kozmová
linguists until these days. I assume, that the relation among valence and intention was
expressed most succinctly by Daneš 9 thereby, that he revealed relation of polysemy
among valence and intention. Within components of valence and intention structure
dominates asymmetric dualism of form and meaning, hence polysemy. The same
valence structure is the base for multiple intention structures. So if we understand
valence as a complex grammatical (not only morpho-syntactic phenomenon), than
intention is the part of valence. That valence, which is also in the spirit of Helbig’s six
stages model10 capable to include all language layers, beginning with morphological
and ending with pragmatical11 layer. In our model, however, we want to follow also
the most recent results, achieved in the area of semantic verb research, verbal situation and semantic markers arguments especially in works of V. Ágel12 and K.
Welke13.
2 To the composition of contrastive – confrontational model:
Valence model of German verb in its latest form14 connects relatively optimal
morpho-syntactic and semantic elements of the valence. Valence model of Slovak
verbs has mainly lexical-semantic character. It is not the objective of this paper to
point to positives or negatives of both valence dictionaries. To the structure of valence
dictionary of Slovak verbs, Jarmila Panevová15 gave her opinion and she herself suggested functional-generative description of verb valence16. In connection to the model
we will in accordance with German linguistics 17 distinguish concept of sentence formula and sentence model.
Sentence formulas are abstract categories, whose amendments can be realized
with various participants. Sentence model is concrete, referring to particular sentence
structure. Even in this regard there are many differences among compared languages.In this way, for example, it is possible in one language to change the subject to
the lateral sentence, in other language; however, the equivalent of verb does not allow
it. The other fact also needs to be stated and that is, that sometimes does not exist
accurate, by convention given translation, equivalents of given verb, actually that for
given verb of initial language exists several equivalents in target language18.
9
F. Daneš / M. Grepl/Z. Hlavsa (1987: 34): Mluvnice češtiny. Praha: Academia.
G. Helbig (1992): Probleme der Valenz- und Kasustheorie. Tübingen: Max Niemeyer.
11
We understand the relation of systemic linguistics and pragmatics as a mutually overlapping,
see also P. Ernst (2004): Einfúhrung in die germanistische Sprachwissenschaft. Wien.
12
V. Ágel (2000): Valenztheorie. Tübingen: Günther Narr.
13
K. Welke (2005): Deutsche Syntax funktional. Tübingen: Niemeyer.
14
H. Schumacher et al. (2004): VALBU Valenzwörterbuch deutscher Verben.
15
J. Panevová (2005): Sloveso: centrum vety, valence: centrálny pojem syntaxe. In: Aktuálne
otázky súčasnej syntaxe, p. 73-74. Bratislava: SAV VEDA.
16
J. Panevová 1974, 1975, 1980, 1994.
17
See especially U. Engel (2004): Deutsche Grammatik. Neubearbeitung. München: Iudicium.
10
Methodological Foundations for Contrastive Model of Verb Valence
225
In both languages verb has also prepositions with the contrary meaning or German
has alternative prepositions (erzählen: von, über) which Slovak does not have. Beside
of that, in Slovak are profusely used so called disposal constructions that present German language does not know. In both languages exists verbs, of which it is possible
under certain circumstances to leave out Esit/Dsit. In such cases is position of stipulation known or comprehended in the frame of context. In most cases, however,
Esit/Dsit realize: goes mainly for verbs: stattfinden/ take place, leben/live, wohnen/live in, sitzen/sit etc.
These and other differences does not allow to create single model of verb in both
languages also considering the fact, that there is no metalanguage, which would
provide objective and complex tertium comparationis. I assume, that also for this
reason most of bilingual valence dictionaries limit itself to use German as initial language and in mother tongue are presented standard equivalents 19of respective German
sentence models.
2.1 Sphere of terminology
In connection to theoretical basis of valence and intention of verb as its part, it is possible to determine also unified terminology for both languages. If we will come out
from the premise, that both German and Slovak linguistics consider as the base of
sentence the sentence formula, resp. particular sentence model, symbols of sentence
complements can be used for both languages. I decided for symbols listed in the latest
work of U. Engl20, therefore we will distinguish among complements subjective:
Esubj/Dsubj,objective namely in accusative: Eakk/Dakk, genitive: Egen/Dgen, dative:
Edat/Ddat, in prepositional case Epräp/Dprep, local: Esit/Dsit, directive: Edir/Ddir or
directional, predicative: Epräd/Dpred and Eexp/Dexp a Everb/Dverb.
In case of adverbial complement, which are being under certain conditions mandatory, e.g. causative, temporal and modal we will use symbol Eadv/Dadv. Instead of
term complement it could be possible to use term complements and supplements. I do
not like this alternative for two reasons. First from didactic reasons term complement
= Ergänzung is commonly used in German language schoolbooks not only on secondary level schools21, but already also on the elementary schools and second even from
linguistic point of view this term is transparent. Subject and object are sentence components not only in the frame of so called traditional syntax, but they have to be
supplemented in the sentence so that it is meaningful, and so are complements becoming subjective or objective comprehensible and usable no matter what is the syntactic
18
R. Kozmová (2007): Zum Problem des kontrastiven Valenzmodells am Beispiel der beVerben. Wrocław. About to be published.
19
As an exception can be considered sentence formulas listed in German-polish grammar:
U. Engel / D. Rytel et al (1999): Deutsch-polnische kontrastive Grammatik. Heidelberg:
J. Groos.
20
U. Engel (2004): Deutsche Grammatik. Neubearbeitung. München: Iudicium.
21
G. Neuner et al.: Deutsch aktiv, Deutsch aktiv neu.
226
Ružena Kozmová
theory, moreover in Slovak linguistics22, term complement is used for nominal predicative complements.
2.2 Categorical meaning
Categorical significance entitles semantic minimum of context components. The minimum requirements for these components are known23, because long time ago24,
before J. Katz a J. Fodor formulated inventory of semantic markers in their work, they
were used in dictionaries and in parallel in many others lexicographic works. There is
a well known fact, that the choice of verb conditions selection of substantive, therefore it is logical, that interest for semantic characteristics of verbs and substantives
was raising together with theoretical development of valence in Anglo-Saxon, as well
as German and Slavonic linguistics.
Concerning analyzed languages, it is necessary to mention works of G. Helbig,
who coming out mainly from theory of deep cases of Ch. Filmore, elaborated relatively in most detailed scope of semantic cases25 in German language. Despite of
certain differences in German and Slovak literature, might be used inventory of
semantic markers of substantives defined by Ulrich Engl26, namely in the widest sense
of the word, because we identify with the above mentioned author in that hierarchy of
semantic markers, videlicet not only substantives, is relatively difficult. The more precise we want to define semantic markers, the more we hit on quantity of not always
secondary problems. On this account, but mainly because, as we deal here with two
different languages, we will not continue as Sokolová in VSS27.
2.2.1 Interdependency of semantic markers of verb and substantive
The choice of functional-semantic method in the analyzing of grammatical phenomenon was in many cases shown as optimal and that is also in the case of the verb
valence. Its ability to bind to itself certain complements is nothing else than ability of
one syntactical unit, by way of semantic markers, form meaningful syntagma on the
basis of compatible markers of other syntactical unit of the same or other word class.
22
See M. Sokolová 1998: 46.
J. A. Katz / J. J. Fodor (1963): The Struture of a Semantics Theory.In: Language 39. p.170210.
24
W. Porzig (1943): Wesenhafte Bedeutungsbeziehungen. In: Beiträge zur Geschichte der
deuschen Sprache und Literatur. p.70-97.
25
In his most recent grammarG. Helbig presents 20 semantic functions: G. Helbig (2002: 468):
Deutsche Grammatik. Ein Handbuch für Ausländer. Berlin / München / Wien / Zürich / New
York: Langenscheidt
26
U. Engel (2004: 188): Deutsche Grammatik. Neubearbeitung. Múnchen: Iudicium
27
J. Nižníková /M. Sokolová (1998): Valenčný slovník slovenských slovies. Prešov:
Slovacontact.
23
Methodological Foundations for Contrastive Model of Verb Valence
227
We will come out from the fact, that verb has inherence meaning, by means of
which it binds with other word classes on the base of their categorical meanings 28.
Following it is combinational characteristics of verb, because it allows binding with
participants and since they enter into relations, it needs to talk also about relational
meaning. All these characteristics of verb determine relation of interdependence of
semantic markers of individual participants and semantic markers of verb, and therefore they participate on the creation of sentence formula, which has its specific
utilized shape in the sentence model. Even the smallest components, semems are
expression of asymmetric form’s dualism and function, because only one sentence
formula can have various semantic partners – participants. It does not have to be
always only agens (Peter cooks a soup), as well as it can be a carrier of process (The
soup is being cooked) or carrier of state (The soup is cooked). Even agens can have
by means of diversely ordered semantic markers of verb and substantive various
modifications. German and Slovak language as accusative languages have more
achievements and accomplishment verbs as verbs representing process. Therefore it is
logical, that achievements verbs are becoming process verbs apart that the syntactical
structure would change. In the spirit of prototypes theory Welke29 explains it in following way:
Prototypic situation has human agens [+hum] as left participant, which deliberately (with own will) in the spirit of intention [+intent] on the base of own activity
[+control] evokes situation, maintains it, suspense and is at the same time responsible
for it [+responsible], thus for its ending, preservation or suspension. Prototypic effect
according to Welke resides in that in case of loss of some marker, e.g. adjacent
remains, therefore existence of agens is only weakened (metaphorically), but agens
even though remains, as it is confirmed by sentence 1, 2.
(1) Petra cooks a good soup.
(2) Our kitchen cooks a good soup.
2.2.2 Verb, verb situation, predicates and actionality in language
In connection to verb situation, resp. with actionality we need to keep in mind great
differences in both analyzed languages. Slovak language is aspectual; aspect is in this
language grammatical category that influences the whole verbal system. German language does not know such a category, and so duration, or one of phases of verb
situation is usually created analytically. Vague semantic even of so called true perfective verbs is more or less evidently demonstrated in equivalents of German
lexeme30. Lemma raten/radiť usually represents imperfect verb situation, but on the
examples 3a, b we can see, that the same sentence can be expressed as perfective as
28
For questions in semantic discipline see also works of J. Kačala (1989): Sloveso
a sémantická štruktúra vety, M. Sokolová: (1993): Sémantika slovesa a slovesný rod.
Bratislava: VEDA.
29
K. Welke (2005: 216 ): Deutsche Syntax funktional. Tübingen: Niemeyer
228
Ružena Kozmová
well as imperfective aspect. And one more example:Verb merken (4) as well as
bemerken (5) has always one equivalent, and that is perfective verb zbadať.
(3) Ich rate dem Kunden, ein anderes Geschäft aufzusuchen.
(3a) Poradím zákazníkovi, aby vyhľadal iný obchod.
(3b) Radím zákazníkovi, aby vyhľadal iný obchod.
(4) Inzwischen hat auch er gemerkt, dass es so nicht geht.
(4a) Medzitým zbadal, (uvedomil si), že to takto nejde.
(5) Das habe ich [gar nicht] bemerkt.
(5a) To som [vôbec] nezbadal (nespozoroval).
In the summary we present sentence models of verb sagen, to which corresponds
Slovak equivalents perfective verb povedať and imperfective verb hovoriť, even
though to Slovak lexeme hovoriť corresponds German lexeme sprechen. This fact
only confirms our previous assertion about vagueness of German verb.
Sentence models: sagen/povedať (say)
a) meaning of verbs sagen, povedať
Meaning
äußern, aussprechen
benutzen, gebrauchen (ein
Wort)
bezeichnen
bei jm. etwas bewirken
informieren (über eine E.)
vorschreiben, befehlen:
vermitteln, ausdrücken, aussagen, berichten
bemerken, feststellen
meinen, behaupten, M. ver30
Example sentences
Er hat mit anderen Worten
das Gleiche wie der Vorredner gesagt.
Sagt man „über jemanden“
oder „von jemandem“
sprechen.
Der Fachmann sagt zum
Schraubenzieher Schraubendreher.
Eine innere Stimme sagt
ihm, dass alles gut sein
wird.
Das haben Sie schön gesagt.
Du hast mir gar nichts zu
sagen!
Niemand kann wirklich
Verläßliches über die Höhe
der Investionen sagen.
Das hat er tatsächlich zu mir
gesagt!
Was sagt denn dein Vater
Equivalent in Slovak
povedať
hovoriť/sagen
(ne)povedať
(na)povedať, našepkať
povedať
hovoriť
povedať
povedať
hovoriť
For aspect questions in connection with Slovak equivalents of German lexeme look into
work of:R.Kozmová: Zum Problem des kontrastiven Valenzmodells am Beispiel der beVerben. Paper on international conference: „Germanistische Linguistik extra muros“ Inspirationen, Aufgaben, Aufforderungen.Wroclav 2008.
Methodological Foundations for Contrastive Model of Verb Valence
treten
argumentieren
annehmen, glauben :
formulieren
zeigen
sich denken, sich überlegen
schweigen
beurteilen
zum Inhalt haben
als Schluss zulassen, besagen, heißen
(Worte, Äußerungen) an
jmdn. richten
erzáhlen mit Dat
sich entfremden
dazu , dass du schon
rauchst?
Dagegen ist nichts zu sagen.
Ich sage, es gibt heute noch
Regen.
Das kann man einfacher
sagen
Der Film sagt mit eindrucksvollen Bildern: Krieg
bedeutet Tod.
Da hab ich mir gesagt: Was
dem einen recht ist, ist dem
andern billig.
Er sagte kein Wort.
Ich kann über dieses Bildnichts Positives sagen.
Das Gesetz sagt eindeutig,
dass ...
Das allein sagt noch nicht
viel.
229
povedať, namietať
myslieť si, domnievať sa
povedať
ukazovať
(po)myslieť, uvedomiť si
mlčať
povedať, posúdiť
hovoriť
(ne)hovoriť
Er sagte ein paar aufmunternde Worte und
verschwand.
Ich habe mir sagen lassen.
Sie haben sich nichts mehr
zu sagen.
povedať
Example sentences
Equivalent in German
Povedal, že sa nemáme znepokojovať
Čo mu na to povedia?
sagen
Anička povedala básničku.
Poviem ti veľkú novinu.
Predavač povedal cenu.
Povedali si, že sa stretnú
pred kinom.
sagen
sagen
sagen
(sich) sagen/sich verabreden
Odborník nepovie šrobovák
ale skrutkovač.
Povedal, že príde.
Táto teória nám povedala
veľa.
sagen
dať si porozprávať
povedať
Povedať:
Meaning of the verb
povedať
vyjadriť slovami, rečou:
vysloviť názor, zaujať stanovisko
predniesť
prezradiť
určiť, stanoviť
povedať si, dohovoriť sa,
dohodnúť sa
pomenovať, nazvať
sľúbiť
obohatiť poznatkami, vysvetliť:
sagen
sagen
(be)sagen
230
Ružena Kozmová
hovoriť pre seba, opakovať si
žalovať,obviniť niekoho z
niečoho
rozhodnúť sa
uznať
niečo namáhavo presadzovať
spôsob formulácie
Povedala si, čo sa naučilla.
Povedal to na mňa otcovi.
(sich) zu sich sagen
petzen
Povedz posledné slovo!
Daj si povedať a poď s nami!
Povedz pravdu, prebiješ si
hlavu.
To ste pekne povedali.
sagen
sagen
sagen
b) generally formula of verb sagen
Verb formula (VF): Esubj, Eakk/Epräp = von + Dat, über + Akk, zu + Dat, (Edat),
Eadv
Semantic marker (SM): sagen/povedať
Esubj = [+control], [+intent], [+reversible], [+hum], [+konkr], [+immat], [intell]
Eakk= [+sachv], [mas/indefinit], [+immat], [+intell]
Edat = [+hum], [+inst]
Epräp = [+hum], [+sachverh]
Eadv/mod = {sonst], [+qual}
NS = dass-S, w-S, ob-S, uNS
c) sentence models:
E = D = complement/doplnenie
NS = VV = subordinative sentence
sentence model 1
Esubj, NS /Esubj, (Korr),
NS
Esubj, NS/HS
Ich sage, es gibt heute noch
Regen.
Myslím si, že dnes bude ešte
pršať.
annehmen, glauben,
meinen, behaupten, M. vertreten,
ukázať, zum Inhalt haben
sentence model 2
Esubj, Eakk
Er hat [mit anderen Worten]
das Gleiche wie der Vorredner gesagt.Povedal [inými
slovami] to isté, čo jeho
predrečník.
äußern, aussprechen, benutzen, gebrauchen (ein Wort),
formulieren,
schweigen (nicht sagen),
Worte, Äußerungen an
jemanden richten
sentence model 3
Esubj, Edat
Da hab ich mir gesagt: Was
dem einen recht ist, ist dem
andern billig.
A tak som si povedal: Čo
jednému vyhovuje, to
druhému vadí.
sich denken, sich überlegen
Methodological Foundations for Contrastive Model of Verb Valence
231
sentence model 4
Esubj, Epräp
Dagegen ist nichts zu sagen.
Voči tomu nemožno nič
namietať.
argumentieren, (nicht) einwenden
sentence model 5
Esubj, Eakk
Das allein sagt noch nicht
viel.
To samo osebe ešte nič neznamená
als Schluss zulassen, besagen,
heißen
sentence model 6
Esubj, Eakk, Epräp/(Epräp)
Der Fachmann sagt zum
Schraubenzieher Schraubendreher. Odborník
nepovie šróbovák, ale
skrutkovač.
vermitteln, ausdrücken, aussagen, berichten, bemerken,
feststellen
sentencemodel 7
Esubj, Eakk, Edat
Du hast mir gar nichts zu
sagen!
Ty mi nemáščo hovoriť
(rozkazovať).
vorschreiben, befehlen
sentence model 8
Esubj, Edat, NS
Eine innere Stimme sagt
ihm, dass alles gut sein
wird. Vnútorný hlas mu
našepkáva, že všetko bude
dobré.
bei jemandem etwas
bewirken.
sentence model 9
Esubj, Eakk, Eadv
Das haben Sie [schön]
gesagt.
To ste pekne povedali
informieren(über eine
Eigenschaft)
sentence model 10
Esubj, Edat, Everb
Ich habe mir sagen lassen.
Dal som si porozprávať.
erzählen mit Dat
c) generally formula povedať
Verb formula (VF): Esubj, Eakk/Epräp =na + Akk , Edat/(Edat), Eadv/mod
Semantic marker (SM): sagen = povedať
VV = že, aby, ako, aký, kedy, či, čo, prečo, s kým
sentence model 1
Dsubj, Dakk
Anička povedala básničku.
Anička trug ein Gedicht
vor.
predniesť, povedať, stanoviť,
rozhodnúť sa, niečo namáhavopresadzovať, pomenovať,
nazvať
sentence model 2
Dsubj, Ddat
Čo mu na to povedia?
Was sagen Sie ihm dazu?
vysloviť názor, zaujať stanovisko
232
Ružena Kozmová
sentence model 3
Dsubj, Dprep
Daj si povedať a poď s nami!
Lass dich überzeugen und
komm mit!
uznať
sentence model 4
Dsubj, VV
Povedal, že sa nemáme
znepokojovať
Er sagte, wir sollten uns
nicht beunruhigen.
vyjadriť slovami, rečou
sentence model 5
Dsubj, Dexp
Táto teória nám povedala
veľa.
Diese Theorie sagte uns
sehr viel.
sentence model 6
Dsubj, Dakk, Ddat
Poviem ti veľkú novinu.
Ich sage dir eine große
Neuheit.
prezradiť
sentence model 7
Dsubj, Ddat, Dprep
Povedal to na mňa otcovi.
Er petzte bei dem Vater
über mich.
žalovať, obviniť niekoho
z niečoho
sentence model 8
Dsubj, Dakk, Dadv
To ste pekne povedali.
Das haben Sie schön gesagt.
spôsob formulácie
sentence model 10
Dsubj, Ddat, VV
Povedali si, že sa stretnú
pred kinom.
Sie haben sich verabredet,
dass sie sich vor dem Kino
treffen.
povedať si, dohovoriť sa,
dohodnúť sa
obohatiť poznatkami, vysvetliť
References
Ágel, Vilmos (2000): Valenztheorie. Tübingen: Narr
Daneš, František / Zdenek Hlavsa(1981): Větné vzorce v češtině. Praha: VEDA
Dolník, Juraj (1990): Lexikálna sémantika: Bratislava: FF UK.
Drosdowski, Günther et al. (1989): Deutsches Universalwörterbuch A. Z.
Mannheim/Wien/Zürich:DUDEN © Duden - Deutsches Universalwörterbuch, 5.
Aufl. Mannheim 2003 [CD-ROM]
Ďurčo, Peter (2009): Modellbeschreibung für Kollokationen mit Basiswort Verb.
S.67-79. In: Deutsche Sprache in der Slowakei.Festschrift fúr Prof. Dr. Ilpo
Tappani Piirainen zum 65.Geburtstag. Internationale Fachtagung Piešťany 13.-15.
Juni 2007.
Methodological Foundations for Contrastive Model of Verb Valence
233
Engel, Ulrich (1976): Liste deutscher Verben mit ihren Bedeutungsbeschreibungen.
In: Engel, Ulrich /Helmut Schumacher u.a.:Kleines Valenzlexikon deutscher
Verben. (Forschungsberichte des Instituts für deutsche Sprache 31). Tübingen:
Narr. Engel, Ulrich (1988): Deutsche Grammatik. Heidelberg: J. Groos.
Engel, U./D.Rytel et al. (1999): Deutsch- polnische kontrastive Grammatik. Bd. 1.
Heidelberg: J. Groos.
Engel, Ulrich (2004): Deutsche Grammatik. Neuberbeitung. München: Iudicium.
Engel, Ulrich/Meike Meliss. Hrsg. (2004): Dependenz, Valenz und Wortstellung.
München: IUDICIUM.
Helbig, Gerhard/M. D. Stepanova (1978): Wortarten und das Problem der Valenz in
der deutschen Gegenwartssprache. Leipzig: VEB Bibliographisches Institut.
Helbig, Gerhard /W. Schenkel, Wolfgang (1983): Wörterbuch zur Valenz und
Distribution deutscher Verben.Tübingen: Max Niemeyer
Helbig, Gerhard (1992): Probleme der Valenz- und Kasustheorie.
Helbig, Gerhard, Joachim Buscha (2002): Grammatik der deutschen Sprache. Ein
Handbuch für Ausländer. Berlin /München/ Wien/ Zürich/ New York: Langenscheidt.
Kačala, Ján (1989): Sloveso a sémantická štruktúra vety. Bratislava: VEDA.
Kačala, Ján (1998): Syntaktický systém jazyka: Pezinok: Formát.
Kačala, Ján et al. (2003): Krátky slovník slovenského jazyka. Bratislava: VEDA.
Katz, Jerry./J. A. Fodor (1963). The structure of a Semantics Theory. In: Language
39. S.170-210.
Lakoff, George/Mark Johnson (1999): Philosophy in the Flesh. In: Computational
Linguistics, vo. 25.
Nižníková, Jolana/Sokolová, Miloslava(1998): Valenčný slovník slovenských slovies.
FF Prešov.
Panevová, Jarmila. (2005): Sloveso:centrum vety, valence: centrálny pojem syntaxe.
73-78. In: Aktuálne otázky súčasnej syntaxe.Bratislava: VEDA SAV.
Pauliny, Eugen (1943): Štruktúra slovenského slovesa. Bratislava: VEDA.
Peciar, Štefan et al. (1959-1968): Slovník slovenského jazyka. Bratislava: VEDA
Schumacher, Helmut./Jacqueline Kubczak/Renate Schmidt, Vera de Ruiter (2004):
VALBU – Valenzwörterbuch deutscher Verben. (Studien zur deutschen Sprache
31). Tübingen: Narr.
Schreiber, Herbert/ Sommerfeldt, Karl-Ernst/Starke, Günter (1993): Deutsche
Substantive. Wortfelder für den Sprachunterricht: Berlin/München u.a:
Langenscheidt – Verlag Enzyklopädie
Sokolová, Miloslava (1993): Sémantika slovesa a slovesný rod. Bratislava: VEDA
Storrer, Angelika (1992): Verbvalenz. Theoretische und methodische Grundlagen
ihrer Beschreibung in Grammatikographie und Lexikograpie. Tübingen: Niemeyer.
Tibenská, Eva (1998): Objektový participant sémantickej štruktúry vety. In:
Slovenská reč, roč. 63, č. 4, s. 198-209.
234
Ružena Kozmová
Tibenská, Eva (2005): Význam a forma v syntaxi. S. 28-36. In: Aktuálne otázky
súčasnej syntaxe. Bratislava: SAV.
Tomášiková, Slavomíra (2008): Ekvivalenty slovenských lokálnych predložiek
v štandardnej a rakúskej nemčine. Acta facultatic philosophicae universitatis.
Prešoviensis. Prešov: FF. Schumacher, Helmut (1986): Verben in Feldern.
Valenzwörterbuch zur Syntax und Semantik deutscher Verben. (=Schriften des
Instituts für deutsche Sprache 1). Berlin/New York: de Gruyter.
Welke, Klaus (2005): Deutsche Syntax funktional, Tübingen: Niemeyer.
Dictionary of Štúr’s Slovak
Ľubomír Kralčák
Faculty of Arts, Constantine the Philosopher University in Nitra, Slovakia
Abstract. Štúr’s Slovak language represents a special landmark signifying
formation of modern literary Slovak language in the history of literary language
of Slovaks. Specific features of Štúr’s language display at every language level,
whereas vocabulary belongs to those levels, which embody the most numerous
representations of the particularities like this.
Dictionary of Štúr’s Slovak Language is a project that shall capture the vocabulary of Štúr’s period as a precious cultural heritage at maximum extent. Hence,
it will be necessary to answer the question of a source dictionary base, i.e. list
of source texts, according to maximum model, in vertical restriction (Štúr’s
texts from the broadest period) and in horizontal restriction (all kinds of texts
from the given period). The project assumes, after setting the source base, the
transfer of the texts into the electronic form, and on this base and by means of
verified software, the creating of lexical data base and its subsequent lexicographical elaboration.
Štúr’s Slovak represents a peculiar landmark in the language history of Slovaks
signifying the beginning of modern standard Slovak. Yet in case of Štúr’s
Slovak it is a concept establishing the structure of modern standard Slovak.
Comparing it to the present-day Slovak, there are considerable differences and
peculiarities which give this language formation a unique form. Specific
features of Štúr’s Slovak show in all the language levels whereby the vocabulary belongs to those subsystems which show the most numerous representation
of such peculiarities. Formation of standard Slovak during Štúr’s period represents simultaneously incredibly vivid, dynamic phase of formation and that not
only from grammatical and syntactical view but also from the view of the range
of functional styles. Lexis has a special status in this context thusly representing
the mentioned unique dynamics most clearly.
1 Štúr’s lexis
This lexis reflects formation processes in a broader sense: it is not only the formation
of the reformulated standard language alone, drafted according to the medieval language structure but also the current social development. Its indicators were new
knowledge, products and social phenomena. Extra language innovations were implemented into the language as an adequate accomplishment of communication needs in
the form of vocabulary innovations.
This process can be first seen in the formation of the so-called cultural lexis which
became the main area of the vocabulary enrichment in the mentioned period. Cultural
lexis is formed by nomination of general abstract concepts, social, cultural, civilizational
236
Ľubomír Kralčák
phenomena and technical terms. The authors of the Štúr’s lexical norm used all the ways
of natural vocabulary enrichment, that is it was about enlarging the meaning of domestic
words, formation of new namings, transfer of foreign words and translation of foreign
words (comp. Horecký, 1946-1948, p. 293; Kondrašov, 1974, p. 179).
On enlarging the meaning, the principle of transfered naming of phenomena was
applied. The most used level of language was spoken one, or more precisely folk
lexis, e.g. kluť sa (“appear, turn up”), roztrhať sa (“to fall apart, to cease to exist by
breaking internal bonds”, e.g. alliance), rozpadať sa (“to split up, to sectionalize into
the several parts”), klesnúť (“to retreat, to become extinct”), viťjecť (“to unfold”) and
others. The same principle was often applied within nouns, e.g. daromňica (“trifle”),
šťjep (“part, section”), štrbina (“blank” abstract meaning), haluz hospodárstva (“sector”), čuridlo (“caricature”), zásťera (“cloak”) and likewise.
Noticeable prevalence of productivity of abstracts is typical mainly of forming
new namings in the area of the so-called cultural lexis, namely expressions generated
by suffix -osť and expressions generated by the sufix -ja/-je. The frequent use of word
formation of the type -osť is evidenced also by using a pronoun as a motivating element if it had an adjective form, e.g. kolkosť (quantity), všetkosť. When it comes to
verbs, it was mainly expressions generated by prefixes. Nearly all the prefixes that
still exist in the contemporary Slovak are held efficient. In Štúr’s Slovak e.g. such prefixed verbs were formed: dotvrdzuvať (something), namožďiť sa (“to worry”),
nadužívať, odmieňať sa (“to take turns”), prečuduvať sa, rozdumať sa, sprofanuvať sa,
vinachádzať sa (“to occur”), zčlánkovať sa (“to unite ”) and likewise.
In terms of suffixes, the most frequently used was -uvať. For instance, premestuvať
sa (somewhere), dozveduvať sa, podsekuvať (something); highly productive was
mailny by forming verbs out of nouns: amalgamizuvať, anatomizuvať, assimiluvať,
kandiduvať and others.
As far as transference of foreign words is concerned, there are borrowed words
mostly from Latin, Greek, less from German and Slavic languages – Czech, Russian,
Polish, Serbo-Croatian language in Štúr’s Slovak (in detail Kondrašov, 1974, p. 209 –
242). In terms of genesis of respective words, we can speak of a specific relationship
between the Czech and Slovak language. According to A. N. Kondrašov (ibid. P. 210)
it is easy to determine the origin of Czech words only within the so-called cultural
words that originated in the first half of the 19th century in Czech and they were then
downloaded to Slovak. In early Štúr’s lexis, there are numerous Czech lexical items,
or more precisely names related to Czech words (e.g. potkať, sklíčiť niekoho “to
deject” , výmluvnosť, dúkaz, rostlina, posvecení and so on.) There is a number of loan
words from Czech in the field of technical terms.
Translation of foreign words was often used to enlarge the number of neologisms
while the original words were written in brackets in order to explain the meaning,
e.g.: Middle Ages (Miitelalter), jet (Einfluss), bill (Wechsel), minority (Minorität),
telegraphy (Telegraph), derivation (Derivatio), fraction (Fractia) and so on.
Dictionary of Štúr’s Slovak
237
Typical feature of Štúr’s lexis was considerable polylexis. Its accompaniment was
a bigger competition among synonymic words, to a great extent among word forming
synonyms. It was usually about the competition among word forming types within
particular word formation categories. For example, this way there were formed synonymous pairs: posilňeňje – posila, vihováraňja – víhovorka, boleňje – boľasť,
zasednuťje – sedňica; by naming the result of the plot: staveňje/stavaňje – stavba (in
more detail in: J. Furdík, 1968), then word forming synonyms were created for
example by naming things: majetok – majetnosť, of things: bažnosť – baživosť, factors
of the plot: spevec – spevák and so on.
As a result of the enlarged lexical variety, new word forming synonymous lines
are being created, and that either as absolute synonyms, e.g. prípad – prípadok –
prípadnosť, or with not absolutely identical semantic features, e.g. cestuvateľ („cestujúci“) – cestovník (e.g. on board ship) – cestujúci (“traveller”). A specific case are
verbs with different preposition phrases, but with the same meaning, e.g. pustiť sa k
niečomu/na niečo, bažiť po niečom/za niečím, ujednostajňiť sa o niečom/na niečom,
zažalosťiť nad niečím/za niečím, opierať sa na niečo/o niečo, natrafiť niečo/na niečo.
Zaznamenali sme aj slovesá s tromi variantnými pádovými väzbami, napr. diviť sa
nad niečím/na niečom/niečomu, oddať sa na niečo/do niečoho/niečomu, zaujímať sa
o niečo/za niečo/niečoho.
Under the given contexts, we can state that lexis of Štúr’s standard Slovak was
characterized by distinct rise of neologisms that appeared throughout the whole journalism of that period. Their number is undoubtedly a sign of a unique extent of
modernization of vocabulary within an uncommonly short time and moreover on a
relatively small text space.
2 The dictionary project
The listed brief characteristics of the lexis of Štúr’s Slovak indicates that its users had
a very inventive, direct approach to the language as a functional and sufficiently shapable means of communication. Štúr’s neologisms represent a peculiar chapter in the
history of standard Slovak, which undoubtedly deserves its own lexicographic processing. Yet if we today look back at the whole period of functioning of Slovak as a
standard, let us say, literary language, we can see that the period of Štúr’s standard
Slovak is least lexicographically processed (a bilingual dictionary by Š. Jančovič:
Noví slovensko-maďarskí a maďarsko-slovenskí slovňík was published in two editions
in 1848). The dictionary of Štúr’s Slovak (DSS) is a long planned project at Nitra’s
Department of Slovak language (comp. Ľ. Kralčák: The project of Štúr’s Slovak and
its computer support, 2001). At the moment this project is becoming more viable,
mainly after becoming a guaranteed VEGA project in 2008.
Already on designing the first draft of the project DSS (comp. Kralčák, ibid.) we
expected three areas of work: collection of language material, lexicographical analysis
of the texts and creation of the lexical database. In the current project there is an
238
Ľubomír Kralčák
ambition to set up at least a part of the core DSS in the work form. Project work is
divided into two stages.
2.1 The main objective of the first stage is to collect language material thusly
making a base for the incoming lexicographical work. This effort has two sides: linguistic and informatic.
2.1.1 The linguistic side od collecting the language material includes selection of
the source base of the dictionary, that is inventory of the source texts. The main aim
of DSS is to record the vocabulary of Štúr’s period as a treasured cultural heritage to
the maximum extent. Therefore it is necessary to solve the question of the source base
of the dictionary according to the “maximalistic model” and that in its vertical delimitation (Štúr’s texts from the broadest period - it is a matter of periodization of Štúr’s
Slovak; in our project this period consists of texts from 1844 to 1852), and in horizontal delimitation (all types of texts from the listed period, hence texts representing
all the developing functional styles). Yet the maximalistic model does not mean that
we will process the complete inventory of the texts of Štúr’s Slovak. After considering
the possibilities of the project team, we made a decision not to include here par excellence texts that were only manuscripts. This does not also include private
correspondence of Štúr’s colleagues but neither texts such as Slovo za slovenčinu by
C. Zoch, drama (fragment of the text) Odchod z Breťislavi by M. Dohnány and others.
The source base is also made by periodicals (Slovenskje pohladi na vedi, umeňja a literatúru, Slovenskje národňje novini, their literary insertion Orol Tatránski, almanach
Ňitra), Štúr’s authorial print, then Francisci, Hodža, Kadavý, Hurban, Dohnány,
Sládkovič (authorial texts are naturally numerous in the above mentioned periodicals).
The listed texts simultaneously represent the most important production from all the
developing functional styles in Štúr’s period.
2.1.2 Informatic side assumes the creation of the lexical database with the help of
the electronic record of the texts. It is the initial step towards the creation of the text
corpus of Štúr’s Slovak. The preparatory project already mentions that there are two
possibilities how to transcribe the texts into the electronic form, namely by the use of
a text editor or a visual scanner of letters. We have opted for the simple transcription
as the visual scanning loses its efficiency due to numerous mistakes which occured
mainly as a result of insufficient quality of the print of the processed material.
2.2 The main objectives for the second stage are as follows:
- lexicographic analysis of the obtained material,
- formation of the corpus of the dictionary of Štúr’s Slovak in the working form.
At this stage we will use the electronic corpus of the texts and by the use of certified software (work with the program WordSmith) we will proceed to the creation of
the lexical database and its follow-on lexicographic processing. With respect to the
fact the work carried out on the project requires a longer lasting research and collecting the language material by more co-workers. The formation alone will be executed
as an output within the stated project only in the working form, or better said depending upon the financial support also in the form of the so-called sample exercise book.
Dictionary of Štúr’s Slovak
239
Other works after completing this stage of the project concentrate on the gradual publishing of the dictionary.
The conclusion. – At this point the project Dictionary of Štúr’s Slovak is being
executed by the team of co-workers at Nitra’s Department of Slovak language. The
project team is presently working on the first stage of the project, that is we focus on
collecting the language data in the electronic form so that we can proceed to its lexicographic processing in the following stage.
References
FURDÍK, Juraj: O tvorení názvov deja a výsledku deja v štúrovskom období. In:
Jazykovedný zborník venovaný VI. slavistickému kongresu. Acta Facultatis
Philosophicae Universitatis Šafarikanae Prešovensis. Bratislava, SPN 1968, s. 89–
106.
HORECKÝ, Ján: K charakteristike štúrovského lexika. Linguistica Slovaca, 4 – 6,
1946 – 1948, s. 279 – 298.
KONDRAŠOV, Nikolaj Andrejevič: Vznik a začiatky spisovnej slovenčiny. Bratislava, Veda 1974. 284 s.
KRALČÁK, Ľubomír: Projekt slovníka štúrovskej slovenčiny a jeho počítačová
podpora. Bratislava: VEDA 2001, s. 150 – 154.
Sources for the Dictionary of Štúr’s Slovak Project
Journals:
Slovenskje pohladi na vedi, umeňja a literatúru. 1846.
Slovenskje pohladi na vedi, umeňja a literatúru. 1847.
Slovenskje pohladi na vedi, umeňja a literatúru. Ďjel I., svazok 3. 1851.
Slovenskje pohladi na vedi, umeňja a literatúru. Ďjel II.Svazok 1- 6. 1851.
Slovenskje pohladi na vedi, umeňja a literatúru Ďjel III. 1852.
Slovenskje pohladi na vedi, umeňja a literatúru Ďjel IV. 1852.
Slovenskje národňje novini.1845, č. 1 - 44.
Slovenskje národňje novini. 1846, č. 45 - 147.
Slovenskje národňje novini. 1847, č. 148 - 248.
Orol Tatránski. Ročník I. 1845/1846.
Orol Tatránski. Ročník II. 1846/1847.
Orol Tatránski. 1848. Ročník III. 1847/1848.
Ňitra. Dar drahím krajanom slovenskím obetuvaní. Ročník II. 1844.
Ňitra. Dar drahím krajanom slovenskím obetuvaní. Ročník III. 1846.
Ňitra. Dar drahím krajanom slovenskím obetuvaní. Ročník IV. 1847.
Novini pre hospodárstvo, remeslo a domáci život. 1848.
240
Ľubomír Kralčák
Other sources:
LICHARD, D.: Domová pokladnica na rok običajní 1847. Skalica. Ročňík prví.
Domová pokladnica na rok prjestupní 1848. Skalica. Ročník druhí.
Domová pokladnica na rok običajní 1849. Skalica. Ročník treťí.
DOHNÁNY, M.: Historia povstaňja z roku 1848. Skalica 1850.
ŠTÚR, Ľ.: Nárečja slovenskuo alebo potreba písaňja v tomto nárečí. V Prešporku
1846. Nauka reči slovenskej. V Prešporku 1846.
FRANCISCI, J.: Slovenskje povesťi. V Levoči 1845.
HODŽA, M. M.: Dobruo slovo Slovákom súcim na slovo. Levoča 1847.
SLÁDKOVIČ, A.: Marína. Bratislava 1846.
HODŽA, M. M.: Ňepi pálenku to je ňezabi. B. Bystrica 1845.
DOHNÁNY, M.: Podmaňínovci. Činohra v 4. ďejstvách. Levoča 1848.
HURBAN, J. M.: Slovo o spolkách mjernosťi a šklách ňeďelních. B. Bystrica 1846.
Českje hlasi proťi slovenčiňe. Skalica 1846.
KADAVÝ, J.: Čítanka pre malje ďjetki. Buďín 1845. Prjaťel ludu. Knižka pre
slovenskích hospodárou a remeselňíkou. Buďín 1845.
JANČOVIČ, Š.: Noví slovensko-maďarskí sa maďarsko-slovenskí slovňík. I. odďjel.
Slovensko-Maďarskí slovňík. II. odďjel. Sarvaš 1848.
Annotation Procedure in Building the Prague
Czech-English Dependency Treebank
Marie Mikulová and Jan Štěpánek
Institute of Formal and Applied Linguistics
Charles University in Prague, Czech Republic
Abstract. In this paper, we present some organizational aspects of building of
a large corpus with rich linguistic annotation, while Prague Czech-English
Dependency Treebank (PCEDT) serves as an example. We stress the necessity to divide the annotation process into several well planed phases. We
present a system of automatic checking of the correctness of the annotation
and describe several ways to measure and evaluate the annotation and annotators (inter-annotator accord, error rate and performance).
1 Introduction
Building a huge corpus with rich linguistic annotation calls for elaborate organization of the annotation process. In our contribution, we will present such a project,
namely Prague Czech-English Dependency Treebank (PCEDT). In the first place,
we will focus on the organizational aspects of the building of the corpus that can be
generally applied to building of any similar huge corpus. In particular, the main
points will be:
– division of the annotation into several phases
– system for checking the accuracy of the annotation
– ways of evaluation of the annotation and annotators
PCEDT is planned to be a corpus of (deeply) syntactically annotated parallel
texts (in English and Czech) intended chiefly for machine translation experiments.
The texts for PCEDT were taken from Penn Treebank [4], which means there are
mostly economical articles from the Wall Street Journal. 2312 documents were used
in PCEDT (approximately 49,000 sentences) that are manually annotated with constituent trees in Penn Treebank. For the Czech part of PCEDT, the English texts
were translated into Czech.
As a base of the process of creation of the corpus (hierarchical system of annotation layers, annotation rules) we will use the already accomplished Prague
Dependency Treebank (PDT) 2.0 [1]. While organizing the annotation of PCEDT
(especially its Czech part, which is the main concern of this article), we will prop
ourselves upon multifarious experiences (both positive and negative) gained from
the production of PDT 2.0.
242
Marie Mikulová and Jan Štěpánek
2 Division of the annotation into several phases
If one builds up a corpus in which a rich and complex linguistic information is
attached to the input data (i.e. sentences), according to our experience it is advisable
to divide this process into several partial phases. The question how to divide the
annotation when the information attached is mostly very complex and various phenomena are interconnected remains rather difficult.
Example of such a rich annotation is the tectogrammatical (deep syntactical)
layer of PCEDT (and similarly the same layer of PDT 2.0). In the annotation process, each sentence is assigned a deep syntactical structure (which among others
deals with ellipsis and valency of verbs and nouns); each unit of the structure is
assigned its deep syntactical function (there are several tens of the “functors”) and
many attributes, mainly grammatemes (tectogrammatical counterparts of morphological categories). The tectogrammatical tree captures coreference, topic-focus
articulation and deep word order. There are 39 different attributes, for a node of a
tectogrammatical tree in PDT 2.0 there are 8.42 attributes filled on average.
The annotation of the tectogrammatical layer was divided into several phases
when the layer was being created for the PDT 2.0 already. The division is inevitable
because no annotator is able to keep all the annotation rules for all the annotated
phenomena in his or her head (the annotation manual [3] has more than 1000 pages).
Moreover, the more information the annotator has to attach to the data, the more
likely he or she omits some of the details.
The experience from production of PDT 2.0 unveiled that the division of the
annotation process into several steps is desirable for the quality of the output data,
even if some phenomena had to be reconsidered repeatedly by different annotators
in various phases.
Today, when building PCEDT, we increased the number of phases even more.
The first phase of the PDT 2.0 annotation (creating a syntactical structure and
assigning functions to the units of the structure) proved to be still too complex and
comprising too many features. We also tried to get rid of repeated resolution of the
same problems, e.g. by introducing new temporary values for some attributes whose
final values will be judged in a later phase of the annotation.
For the tectogrammatical annotation, we count on the following phases:
1. building a tree structure, dealing with ellipsis included; assignment of functors
and valency frames, links to lower layers (10 attributes).
2. annotation of subfunctors (fine grained classification of functors, 1 attribute).
3. annotation of coreference (4 attributes).
4. annotation of topic-focus articulation, rhematizers and deep word order (3 attributes).
5. annotation of grammatemes, final form of tectogrammatical lemmata (17 attributes).
6. annotation of remaining phenomena (quotation, named entities etc.)
Annotation Procedure in Building the Prague Czech-English Dependency Treebank
243
The first phase is still the most difficult, each annotator is responsible for the
whole structure of the tree and correct values of ten attributes. All these attributes
are connected with the structure and deep syntactical functions of the nodes. The
annotator does not have to pay attention to anything else.
For the first phase of the annotation process a “working value” #NewNode was
established for tectogrammatical lemmata of nodes added to trees in case of ellipsis
of valency frame arguments. Absent obligatory arguments are represented by added
nodes in final tectogrammatical trees and their tectogrammatical lemmata signify
the type of the elision (#Gen stands for a general participant, #PersPron for a deletion, #Cor and #QCor for a controlee in control constructions, #Rcp for ellipses
because of reciprocation). The type of elision is closely connected with coreference
(some types of absent arguments have a coreferent, some do not). During the annotation of PDT 2.0, the lemmata of absent arguments were assigned in the first phase,
the coreference in a following phase. By introducing the #NewNode value the final
solution of the tectogrammatical lemma was postponed to the following phase
together with remaining questions of coreference. An annotator inserts a node with
the “working value” of the tectogrammatical lemma and only assigns its syntactical
function, not taking care about the lemma.
In the first phase we also “neglect” the annotation of rhematizers (in PDT 2.0,
they were annotated in the first phase). Competent decision about a rhematizer
(whether an expression is a rhematizer or not, its position in the tree) is possible
only if the topic-focus articulation of the sentence is decided at the same time.
Therefore, definitive annotation of rhematizers is planned to the topic-focus annotation phase.
Determine an amount of annotated information that does not harm the quality of
the data is obviously difficult. Our believe that the current schedule of phases constitutes a reasonable and manageable rate seems to be justified by the measuring of
inter-annotator agreement (see section 4.1). The quality of the data is regularly
guarded by a system of automatic checking procedures (see section 3).
3 System for checking the accuracy of the annotation
When PDT 2.0 was in production, only random “manual” checks of the accuracy
were performed. The real checking took place when all the annotation had finished.
The checking and fixing phase was quite complex and time-consuming; moreover,
in some cases, the changes were not realized full-scale [5].
We want to avoid such a procedure in the development of PCEDT. Checking of
the data is performed in parallel with the annotation process. At the beginning of the
process, a number of automatic checking procedures was proposed and new tests
subsequently come up during the annotation process. Currently there are 99 checking procedures that verify the annotation of Czech sentences.
The checking procedure proposal is based on the fact that many annotation rules
imply that particular phenomena cannot (or have to) occur in the annotation output.
They mainly combine attribute values and structure of a tree. For example, a simple
244
Marie Mikulová and Jan Štěpánek
check states that every coordination has at least two members and reports all onemember coordinations as errors. Another check states that the root of a tectogrammatical tree has only a limited set of possible functors (PRED for a predicate,
DENOM for nominative clause, PARTL for interjection clause etc.). There is also a
converse check monitoring that no dependent node has the PRED functor, and so on.
The checking procedures return a list of erroneous (questionable) positions in the
data. The annotator gets his or her data back for corrections, manually fixing each
position.
The checking procedures are run periodically after a given volume of the data
has been annotated (1000 sentences) or once a quarter. All the data are checked
every time (in case a new check existed) and after the correction, the data are
checked again and again while there are any errors (new errors can arise in fixing
the old ones).
Automatic checking procedures improve the quality of the data not only by fixing the present errors, but also by providing a feedback to the annotators (because
each annotator fixes his or her own data, i.e. his or her own errors) and thus eventually improving the future annotation.
4 Ways of evaluation of the annotators
A system for evaluation of the annotation and annotators should by an integral part
of any annotation project. In the PCEDT project, the quality of the work of a particular annotator is judged by several ways:
• the annotation agreement between annotators is measured,
• the output of the automatic checking procedures tells us how often an annotator
makes mistakes compared to the others,
• the annotators book the time they spend annotating; it allows later to evaluate
their performance and the relation of the efficiency to the error rate.
4.1 The annotation agreement between annotators
The basic way how to evaluate an annotation is to measure the inter-annotator
agreement. However, the structure to be compared is very complex. The algorithm
aligning two tectogrammatical trees built upon the same analytical tree is described
in detail in [3], once the trees are aligned node to node, we just compare the values
of all the attributes of all the aligned nodes. To evaluate the structural agreement,
we treat the identifier of a node's parent as a new attribute of the node. Complex
attributes (lists, structures etc.) need further manipulation in order to be compared.
For example, identifiers of linked analytical nodes have to be sorted; for annotator's
comment, we only compare the type, because the text can vary.
Since there is no “golden” annotation, we just measure the agreement of all the
pairs of annotators (see Table 1, data from December 2007; average value is shown
for every attribute, and average value over all the attributes and structure is
presented as “Overall”). As a baseline, we use the output of an automatic procedure
with which the annotators start their work (marked “Z” in the table). Note that the
Annotation Procedure in Building the Prague Czech-English Dependency Treebank
245
agreement among annotators is always higher than the agreement between any
annotator and the baseline. The attributes with a lower difference between baseline
and the annotators (about 5%, i.e. is_state, is_generated, is_dsp_root, compl.rf,
annot_comment, and a/lex.rf) tend to contain more errors, or have too vague
annotation rules.
The annotator that agrees most with all the others (“K”) is at the same time the
annotator that makes the least errors and submits the most sentences (see next
sections).
Overall
Structure
a/aux.rf
a/lex.rf
annot_comment
compl.rf
functor
K
Ma
A
O
Z
A
Ma
O
K
Z
K
Ma
A
O
Z
K
Ma
A
O
Z
K
Ma
A
O
Z
K
Ma
A
O
Z
K
Ma
O
A
Z
94,08%
94,01%
93,83%
93,78%
84,58%
88,62% is_dsp_root
88,60%
87,92%
87,88%
69,28%
93,82% is_generated
93,58%
93,55%
93,53%
82,45%
96,26% is_member
96,12%
96,00%
95,90%
89,67%
96,52% is_parenthesis
96,40%
96,30%
96,27%
90,43%
96,32% is_state
96,22%
96,12%
96,03%
90,18%
85,70% t_lemma
85,67%
85,57%
85,13%
66,80%
K
A
Ma
O
Z
K
A
Ma
O
Z
K
A
Ma
O
Z
Ma
K
O
A
Z
K
Ma
O
A
Z
K
Ma
O
A
Z
Table 1. Inter-annotator agreement
95,86%
95,83%
95,75%
95,72%
89,72%
96,24%
96,05%
96,03%
96,02%
90,27%
94,72%
94,70%
94,50%
94,25%
85,47%
95,42%
95,40%
95,27%
95,15%
88,72%
96,50%
96,25%
96,13%
96,13%
90,35%
93,76%
93,60%
92,70%
92,42%
81,60%
246
Marie Mikulová and Jan Štěpánek
4.2 Error rate
Using the list of errors generated by the checking procedures we can count how
often the annotators make errors (only those errors the procedures can detect, of
course): the number of errors the annotator made is divided by the number of
sentences or nodes she annotated. Table 2 shows the comparison of the error rate for
4 annotators in December 2007 (at the beginning of the process) and current
numbers for 7 annotators from July 2009. The numbers from different periods
cannot be compared directly because since the beginning there have been more than
30 new checking procedures, which means the current list of errors is longer. On the
other hand, the rank of the annotators can be compared.
The table shows that our current best annotator (“K”) had approximately 30
errors per 100 sentences and 1.62 errors per 100 nodes. Her error rate has not got
worse over the two years and she remains the best annotator. The table further
shows that the differences in error rate between annotators can be great and that all
the annotators keep their positions: no one gets markedly better nor worse. The comparison of veteran annotators and the new ones that annotate only for a short time is
also interesting: it shows that knack, practice, and experience lead to quality of the
annotation.
December 2007
July 2009
Who
Errors per 100
sentences
Errors per 100
nodes
Errors per 100
sentences
Errors per
100 nodes
K
29.7851
1.6241
1.5103
0.0806
O
39.6699
2.0624
4.0331
0.2067
Ma
61.4087
3.2707
8.4670
0.4533
A
63.2318
3.3498
6.3583
0.3265
L
-
-
15.0668
0.8010
Mi
-
-
16.2241
0.8460
J
-
-
19.0476
1.0971
Table 2. Error rate
4.3 Performance of the annotators
In the annotation process, even the time spent working by the annotators is measured. The annotators book the time to a web form. For each month the web
application counts the annotators' performance over the month and the over-all performance. The data are important among others to determine the wages; on the basis
of the data we tariff a sentence (annotators are being paid monthly according to the
number of sentences they have annotated).
Annotation Procedure in Building the Prague Czech-English Dependency Treebank
247
Table 3 shows performance of the annotators in October 2008 and June 2009,
table 4 shows the over-all performance. Monitoring the performance illustrates the
differences between annotators, but also the fluctuation of each particular annotator.
We can also observe the inverse proportionality of the performance and error rate
(see section 4.2): the more is the annotator efficient (she annotates more data), the
less errors she makes.
October 2008
Who
June 2009
Sentences Minutes per
Hours Sentences
Hours
per hour sentence
Sentences
Sentences Minutes per
per hour sentence
A
18.50
147
7.946
7.551
-
-
-
-
I
100.50
742
7.383
8.127 101.50
1229
12.108
4.955
J
11.50
97
8.435
7.113
2.00
28
14.000
4.286
K
33.00
418
12.667
4.737
23.50
332
14.128
4.247
L
46.00
143
3.109
19.301
27.88
365
13.092
4.583
Ma
40.00
310
7.750
7.742
-
-
-
-
Mi
17.85
142
7.955
7.542
24.91
358
14.372
4.175
O
37.81
403
10.659
5.629
56.65
632
11.156
5.378
Table 3. Performance of the annotators
Who Hours
Sentences
Sentences per hour
Minutes per sentence
A
114.25
963
8.4289
7.1184
I
827.00
7006
8.4716
7.0825
J
105.70
1001
9.4702
6.3357
K
107.00
1430
13.3645
4.4895
L
266.41
1716
6.4412
9.3150
Ma
78.00
615
7.8846
7.6098
Mi
169.98
1655
9.7364
6.1624
O
289.02
3211
11.1100
5.4006
Table 4. Over-all performance of the annotators
5 Conclusion
In the article, we have presented some organizational aspects of building of a large
syntactical treebank. We stressed mainly the necessity to divide the annotation process into several well planned phases. We presented out system for checking the
correctness of the annotation. The fact that the correctness is being checked at all
should be pointed out: it is not a common practice in similar projects. We described
three ways to measure and evaluate the annotation and annotators.
248
Marie Mikulová and Jan Štěpánek
We believe that having published PDT 2.0 with 50,000 sentences annotated on
the tectogrammatical layer and being in the halftime of the PCEDT project with
more than a half data already annotated (33,500 sentences, 68% of the corpus) our
proposals are sufficiently backed by our experience and practice.
Acknowledgement:
The research reported in this paper was supported by the LC536, GAUK
22908/2008, and FP6-IST-5-034291-STP.
References
[1] Hajič, J. et al. (2006). The Prague DependencyTreebank 2.0. CD-ROM. Linguistics Data Consortium Cat. No. LDC2006T01. Philadelphia, PA, USA. URL:
http://ldc.upenn.edu, http://ufal.mff.cuni.cz/pdt2.0
[2] Klimeš, V. (2006). Analytical and Tectogrammatical Analysis of a Natural
Language. PhD Thesis. MFF UK, Prague.
[3] Mikulová, M. et al. (2006). Annotation on the tectogrammatical level in
Prague Dependency Treebank. Annotation manual. Technical report ÚFAL
TR-2006-30. MFF UK, Prague.
[4] Mitchell, M. et al. (1995). Treebank 2. CD-ROM.. Linguistics Data
Consortium Cat. No. LDC95T7. Philadelphia, PA, USA. URL:
http://ldc.upenn.edu
http://www.cis.upenn.edu/~treebank/
[5] Štěpánek, J. (2006). Závislostní zachycení větné struktury v anotovaném
syntaktickém korpusu (nástroje pro zajištění konsistence dat) [Capturing a
Sentence Structure by a Dependency Relation in an Annotated Syntactical
Corpus (Tools Guaranteeing Data Consistence)]. PhD Thesis. MFF UK,
Prague.
Automatic Analysis of Terminology
in the Russian Corpus on Corpus Linguistics
Olga Mitrofanova1 and Victor Zakharov1,2
2
1
Saint-Petersburg State University
Institute of Linguistic Studies of the Russian Academy of Sciences, Russia
[email protected]
[email protected]
Abstract. The paper is devoted to processing terminological items occurring in
domain-restricted texts. Linguistic resource involved in research is the Russian
corpus on corpus linguistics developed in St. Petersburg State University and
Institute of Linguistic Studies, RAS. (Semi-)automatic terminology extraction is
performed with the help of linguistic and statistical tools which allow to
generate lists of single-word and multi-word terms supplied with frequency data
and lexical-syntactic patterns. Lexical-syntactic patterns are used in the analysis
of contexts which contain definitions of terms, expose interrelations between
terms, provide their synonyms, translation equivalents, etc. Results obtained
contribute much to the construction of the Russian thesaurus on corpus linguistics.
1 Introduction
Domain-restricted text corpora reflecting knowledge on particular subject areas stand
out against a background of the variety of corpus genres. The given type of corpora is
distinguished by rigid restrictions in kinds and topics of texts, formalization of text
content due to logical and conceptual structure of subject areas; transparent hierarchical structure of the lexicon owing to its saturation with terms; the influence of
scientific content and style on lexical, semantic, morphological, syntactic parameters
of corpus texts [Gerd 2005].
Results of the analysis of corpora developed for particular subject areas are of
great practical importance. Domain-restricted text corpora and linguistic data extracted from them are used both in scientific and technical lexicography (in compiling
terminological dictionaries, classifiers, etc.) and in the field of automatic text processing (automatic indexing, summarization and classification, information retrieval,
machine translation, etc.). Domain-restricted text corpora serve as a basis for
(semi-)automatically created terminological dictionaries and thesauri, formal ontologies, mono- and multilingual resources.
Research in the field of development and further application of domain-restricted
text corpora created for new subject areas attract much attention of linguists and terminologists. Certainly, corpus linguistics itself belongs to such new areas.
Complex description of corpus linguistics terminology has been performed for
English [Baker et al. 2006] as well as for some other languages including Slavic, cf.
250
Olga Mitrofanova and Victor Zakharov
corpus linguistics section of the Slovak terminological database which was built in
Ľ. Štúr Institute of Linguistics, the Slovak Academy of Sciences, Bratislava [Levická
2007; Šimková 2006] (URL: http://data.juls.savba.sk/std/). However,
until recently, this subject area was not presented in Russian terminology.
The main tasks of the project carried out by the Department of Mathematical Linguistics, St. Petersburg State University and Institute of Linguistic Studies of the
Russian Academy of Sciences are the creation of the Russian corpus on corpus linguistics, the Russian thesaurus on corpus linguistics and development of a range of
linguistic resources on its basis.
The present paper discusses particular aspects of the project, namely, automatic
extraction, processing and systematization of terms from the Russian corpus on corpus linguistics.
2 The Russian corpus on corpus linguistics and conjoined
linguistic resources
The Russian corpus on corpus linguistics includes numerous text reflecting a wide
range of corpus linguistics problems, among them definition of corpus linguistics as a
particular sphere of research activity; its opposition to other divisions of linguistics
and language engineering; definition of a corpus with respect to other kinds of linguistic data; procedures of creating and using corpora (tagging, alignment, searching,
etc.); corpus typology; developers and users of corpora; interrelations between corpora and other linguistic resources.
The core part of the Russian corpus on corpus linguistics consists of research
papers in Russian: conference proceedings [Corpora 2002, 2004, 2006, 2008, etc.],
separate articles, manuals, textbooks, monographs, etc. The corpus is regularly
updated and enriched with new materials.
Each text of the corpus is preprocessed: special tags for tables, images, formulae,
links, numbers, non-Russian text fragments, etc. are introduced, morphological tagging and lemmatization are performed. Further the texts are supplied with metadata
(metatagging) which include bibliographic passport and a set of terms (key words)
indicating the topic of the paper.
Cluster analysis of terms (key words) [Mitrofanova, Panicheva, Savitsky 2007]
allowed to form the nuclear part of formal ontology which reflects major concepts of
corpus linguistics [Vinogradova, Mitrofanova 2008]. Expansion of formal ontology
implies automatic retrieval and further analysis of terms occurring in the processed
texts. Among other tasks of the project are thematic classification of texts, construction of the Russian thesaurus on corpus linguistics, etc. To fulfill these tasks it is
necessary to perform (semi-)automatic extraction of terms from the corpus.
Automatic Analysis of Terminology in the Russian Corpus on Corpus Linguistics
251
3 Techniques used for term extraction and Alex+
terminological toolkit
Major approaches to term extraction often require linguistic and/or statistical analysis
of corpora. Linguistic techniques of term extraction imply mainly manual text processing performed by experts who select supposed single-word and multi-word terms.
Statistical techniques treat terms as frequent lexemes or word-groups. Multi-word
terms are regarded as n-grams characterized by high values of association measures
(e.g. MI-score, t-score, Log-Likelihood, C-value, χ2, etc.).
Combination of linguistic and statistical techniques seems to be highly productive
in the analysis of terminology: cf. research results dealing with Slavic languages, Russian in particular [Braslavskij, Sokolov 2008; Dobrov et al. 2003; Kupść 2007;
Urbańska, Piechociński 2007]. Hybrid approach to term extraction deals with
(semi-)automatic processing of domain-restricted corpora and require the use of lexical-syntactic patterns, various filters (e.g. stop-word dictionaries) alongside with
frequency analysis and collocation extraction [Zakharov, Khokhlova 2008].
Various toolkits aimed at (semi-)automatic term extraction have been developed,
among them Alex+ for Russian [Sidorova 2008]. Alex+ turns out to be a convenient
lexicographic environment for development and maintenance of domain-restricted
dictionaries and thesauri which allows to process corpora, extract single-word and
multi-word terms from texts taking into account certain lexical-syntactic patterns, perform statistical analysis of texts in corpora, automatically enrich dictionaries using
learning samples. Alex+ includes modules for morphological analysis, assemblage of
multi-word terms according to particular lexical-syntactic patterns, concordance creation and revision, stop-words detection, etc. Dictionaries developed with the help of
Alex+ include various types of information about terminology (domain data, semantic
features, concept hierarchy, statistical features, etc.). Alex+ allows to develop formal
ontologies alongside with dictionaries and thesauri which can be used simultaneously
for classification of texts in corpora. Thus, possibilities of Alex+ fully correspond to
the tasks of research project in question, that’s why it was used for (semi-)automatic
extraction of terms from the Russian corpus on corpus linguistics.
4 Research results
4.1 Description of single-word and multi-word terms with the help of
lexical-syntactic patterns
Texts included in the Russian corpus on corpus linguistics were processed with the
help of Alex+ so that lists of single-word and multi-word terms were successfully
formed and sorted with regard to general lexical-syntactic patterns, frequency, etc.
It is very important to distinguish carefully terms and non-terms. On the one hand,
one should rely upon expert opinions; on the other hand, there are certain criteria
allowing to solve the problem, e.g. specificity criterion [Šajkevič 2003].
252
Olga Mitrofanova and Victor Zakharov
Lists of terms were edited: stop-words (function words) were marked out and
added to stop-word dictionary, non-terms (e.g. miro) were deleted.
List of single-word terms include nouns, adjectives and verbs, e.g.:
N: vydača, dannyje, dokument, zapros, lemma, metka, razmetka, častota, etc.;
Adj: avtomatizirovannyj, informacionno-poiskovyj, korpusnoj, korpusnyj, , etc.;
V: avtomatizirovat’, razmečat’, etc.
List of multi-word terms include word-groups representing various lexical-syntactic patterns, e.g.:
Adj+N: avtomatičeskaja obrabotka / razmetka / systema, avtomatičeskij analiz /
režym, anaforičeskaja / morfologičeskaja / semantičeskaja / sintaksičeskaja / strukturnaja /prosodičeskaja razmetka, etc.;
N+N: korpus dannych / tekstov, model’ jazyka, razmetka korpusa / dokumenta / teksta, razmer korpusa, raspoznavanije reči, etc.;
N+Prep+N: dostup k korpusu, nauka o jazyke, poisk v korpuse, etc.;
N+Adj+N: bank sintaksičeskich struktur, massiv jazykovych dannych, obrabotka
tipovych zaprosov, etc.;
N+Prep+Adj+N: korpus s sintaksičeskoj razmetkoj, teksty na jestestvennom jazyke,
etc.;
Adj+N+N: avtomatičeskaja obrabotka teksta, kompjuternaja baza dannych, kompjuternaja model’ jazyka, etc.;
N+N+N: vyvod rezul’tatov poiska, standart predstavlenija metadannych, etc.;
N+Prep+N+N: poisk s ukazanijem konteksta, etc.;
Adj+Adj+N: ustnaja razgovornaja reč , etc.
The most frequent lexical-syntactic patterns are Adj+N, N+N, N+Adj+N,
Adj+N+N, N+N+N.
Multi-word terms exposed and described with the help of Alex+ differ not only in
structure and content but also in cohesion, especially tree- and four-word terms, as
they are supposed to be combined of one- and two-word terms. Level of cohesion can
be evaluated in course of statistical analysis.
4.2 Description of terminological contexts with the help of
lexical-syntactic patterns
Lexical-syntactic patterns are successfully used in the analysis of terminological contexts which contain definitions of terms, expose interrelations between terms, provide
their synonyms, translation equivalents, etc. Annotation of terminological contexts
allows to point out special context markers which can be used in selection of terms
from the processed texts and revelation of their significant features. The structure and
typical content of terminological contexts may be described by various lexical-syntactic patterns, e.g.:
Automatic Analysis of Terminology in the Russian Corpus on Corpus Linguistics
253
NP(def) <nazyvat’ / nazyvat’s’a / imet’ nazvanije> NP(term)
(definitions of terms):
Eto kodirovanije informacii imejet nazvanije metarazmetka… [Zakharov 2005];
NP(term) <, ili> NP(term)
(synonymic relations between terms):
…sintaksičeskogo analiza, ili parsinga…[Zakharov 2005];
NP(term) <javl’at’s’a rezul’tatom> NP(term)
(«process – result» relations between terms):
……sintaksičeskaja razmetka, javl’ajuš’ajas’a rezul’tatom sintaksičeskogo analiza, ili
parsinga… [Zakharov 2005];
NP(term) <obespečivat’> NP(term)
(«item – purpose» relations between terms):
…konvertirovanije razmečennych tekstov v strukturu specializirovannoj lingvističeskoj informacionno-poiskovoj sistemy (corpus manager), obespečivajuš’ej
bystryj mnogoaspektnyj poisk i statističeskuju obrabotku … [Zakharov 2005].
Analysis of terminological contexts contributes much to exposure of term hierarchy, specification of the word list for the developed thesaurus, enrichment of the
thesaurus with definitions. Definition module of the thesaurus includes definitions
extracted directly from texts of the corpus and definitions constructed according to
regular lexical-syntactic patterns occurring in the corpus. The use of lexical-syntactic
patterns provides uniformity of descriptions given to terms in the thesaurus. In future
it seems reasonable to use special language for annotation of lexical-syntactic patterns, e.g. LSPL (Lexical-Syntactic Pattern language) [Bolšakova et al. 2007;
Rabčevskij et al. 2008].
5 Conclusion
The given research allowed working out the strategy of (semi-)automatic terminology
extraction from the Russian corpus on corpus linguistics. Alex+ terminological toolkit
was used for single-word and multi-word term extraction. Major lexical-syntactic patterns of single-word and multi-word terms were revealed and described.
Terminological contexts occurring in the corpus were analyzed with the help of
enriched lexical-syntactic patterns as well.
Results of automatic terminology extraction contribute much to the construction of
the Russian thesaurus on corpus linguistics. It is assumed that the developed terminological thesaurus should inherit the hierarchy of concepts reflected in the formal
ontology on corpus linguistics; it will contain terms characteristic of the domain, their
definitions, synonyms and translation equivalents, as well as bibliographic descriptions of texts in which they were registered.
It is expected that the Russian thesaurus on corpus linguistics will be integrated
into the knowledge portal on computational linguistics developed by a group of Russian researches (Moscow, Novosibirsk and St. Petersburg) [Sokolova et al. 2008].
254
Olga Mitrofanova and Victor Zakharov
References
[1] Backer, P.; Hardie, A.; McEnery, T.: A Glossary of Corpus Linguistics.
Edinburgh University Press (2006)
[2] Bol’šakova, E. I.; Bajeva, N. V; Bordačenkova, E. A.; Vasiljeva, N. E.; Morozov, S. S.: Leksiko-sintaksičeskije šablony v zadačach avtomatičeskoj
obrabotki teksta. In: Kompjuternaja lingvistika i intellektual’nyje tehnologii:
Trudy meždunarodnoj konferencii «Dialog–2008». Moscow (2008)
[3] Braslavskij, P. I.; Sokolov, E. A.: Sravnenije p’ati metodov izvlečenija
terminov proizvol’noj dliny. In: Kompjuternaja lingvistika i intellektual’nyje
tehnologii: Trudy meždunarodnoj konferencii «Dialog–2008». Moscow (2008)
[4] Corpora2002 – Doklady naučnoj konferencii «Korpusnaja lingvistika i lingvističeskije bazy dannych – 2002». St. Petersburg (2002)
[5] Corpora2004, 2006, 2008– Trudy meždunarodnoj konferencii «Korpusnaja
lingvistika – 2004 (2006, 2008)». St. Petersburg (2004, 2006, 2008)
[6] Dobrov, B. V.; Lukaševič, N. V.; Syrom’atnikov, S. V.: Formirovanije bazy
terminologičeskich sočetanij po tekstam predmetnoj oblasti. In: Trudy 5j Vserossijskoj konferencii «Elektronnyje biblioteki: perspektivnyje metody i
technologii, elektronnyje kollekcii» RCDL–2003. St. Petersburg (2003)
[7] Gerd, A. S.: Jazyk nauki i techniki kak object lingvističeskogo izučenija. In:
A.S. Gerd. Prikladnaja lingvistika. St. Petersburg (2005)
[8] Kupść, A.: Extraction automatique de termes à partir de textes polonaise. In:
TALN 2007. Toulouse (2007)
[9] Levická, J.: Terminology and Terminological Activities in the Present-Day
Slovakia. In: Computer Treatment of Slavic and East European Languages:
Fourth International Seminar. Bratislava, Slovakia, 25–27 October 2007.
Proceedings. Bratislava (2007)
[10] Mitrofanova, O.; Panicheva, P.; Savitsky V.: Automatic Word Clustering in
Russian Texts based on Latent Semantic Analysis. In: Computer Treatment of
Slavic and East European Languages: Fourth International Seminar.
Bratislava, Slovakia, 25–27 October 2007. Proceedings. Bratislava (2007)
[11] Rabčevskij, E. A.; Bulatova, G. I.; Šarafutdinov, I. M.: Formalizm zapisi
leksiko-sintaksičeskich šablonov v zadače avtomatizacii postrojenija ontologij.
In: Trudy 10j Vserossijskoj konferencii «Elektronnyje biblioteki: perspektivnyje metody i technologii, elektronnyje kollekcii» RCDL–2008. Dubna (2008)
[12] Sidorova, E. A.: Podchod k postrojeniju predmetnych slovarej po korpusu
tekstov In: Trudy meždunarodnoj konferencii «Korpusnaja lingvistika – 2008».
St. Petersburg (2008)
[13] Sokolova, E. G.; Kononenko, I. S.; Zagorulko, Ju. A.: Problemy opisanija
kompjuternoj lingvistiki v vide ontologii dl’a portala znanij. In: Kompjuternaja
lingvistika i intellektual’nyje tehnologii: Trudy meždunarodnoj konferencii
«Dialog–2008». Moscow (2008)
Automatic Analysis of Terminology in the Russian Corpus on Corpus Linguistics
255
[14] Šajkevič, A. Ja.: Statističeskij slovar’ jazyka Dostojevskogo. Vvedenije.
Moscow (2003)
[15] Šimková, M.: Výberový slovník termínov z počítačovej a korpusovej
lingvistiky (2006) URL:
http://korpus.juls.savba.sk/publications/block1/2006
-simkova-vyberovy%20slovnik%20terminov/2006-simkovavyberovy%20slovnik%20terminov.pdf
[16] Urbańska, D.; Piechociński, D.: Automatic Term Recognition in Polish Texts.
In: Computer Treatment of Slavic and East European Languages: Fourth
International Seminar. Bratislava, Slovakia, 25–27 October 2007. Proceedings.
Bratislava (2007)
[17] Vinogradova, N. V.; Mitrofanova, O. A.: Formal’naja ontologija kak instrument sistematizacii dannych v russkojazyčnom korpuse tekstov po korpusnoj
lingvistike. In: Trudy meždunarodnoj konferencii «Korpusnaja lingvistika –
2008». St. Petersburg (2008)
[18] Zakharov, V. P.: Korpusnaja lingvistika / Učebno-metodičeskoje posobije. St.
Petersburg (2005)
[19] Zakharov, V. P.; Khokhlova, M. V.: Statističeskij metod vyjavlenija
kollokacij. In: Jazykovaja inženerija v poiske smyslov. XI Vserossijskaja
objedinennaja konfererncija «Internet i sovremennoje obsčestvo».
St. Petersburg (2008)
Using Speech and Handwriting Recognition in
Electronic School Worksheets
Marek Nagy
Department of Applied Informatics,
Faculty of Mathematics, Physics and Informatics, Comenius University
Mlynská dolina, 842 48 Bratislava, Slovak Republic
[email protected],
http://www.ii.fmph.uniba.sk/~mnagy
Abstract. At elementary schools typical paper worksheets are used. This approach has some disadvantages. A bit clever children fill in it quickly and then
they start to bore although they have done some mistakes. From this point of view
electronic worksheets with an instant feedback will be more appropriate. There
is not problem adopt activities in click-click style but more important activities
have writing, reading and listening style at early grades. In this contribution the
worksheets for Slovak ABC book Lipka were studied and adopted. To make the
instant feedback speech and handwriting recognition is used. An attention will
be focused on an isolated word speech recognition/confirmation where a phonetically based approach is used.
1 Introduction
Children use typical paper worksheets at elementary schools. Many kinds of exercises
can be depicted on the worksheets. Children read a legend of an exercise and fill in it.
This schema is especially problematic for young children who cannot read very well.
Of course, first graders cannot read absolutely. This problem is solved by teachers. The
teacher explains the exercise and all children of class following him. After, children can
start filling in the exercise. It is clearly that this approach needs a synchronization of all
children. But a bit clever children fill in it quickly and then they start to bore although
they have done some mistakes. The teacher has not time to correct all mistakes of children quickly. Electronic worksheets with an instant feedback can be more appropriate
for this purpose.
I will concentrate on first-graders worksheets in this contribution. I analyzed Slovak
ABC book Lipka [3] and an appertaining worksheet [2]. The problem with text legend
which cannot be read by children is partially solved in [2] by a set of pictographs. These
pictographs can help to identify types of exercises and help to manage an adaptation in
an electronic form. One exercise from the worksheet is depicted in Figure 1. Typical
activities are listen to, read aloud/entitle, encircle, write down or color (Figure 2). I divide these activities into two groups: vocal activities and writing activities. The writing
activities use handwriting recognition mainly. Two approaches must be considered. If
children write down letters according to patterns then dynamic programming comparison would be used [1, 7]. On the other hand, if children fill in letters or numbers then
independent recognition which uses statistical approach is better [11].
Using Speech and Handwriting Recognition in Electronic School Worksheets
Fig. 1. An example of an typical exercise from worksheet of ABC book Lipka [2]. The legend
of exercise in English: “Entitle the pictures. Encircle letters whose phones you can hear at the
beginning of words.”
Fig. 2. Some pictographs for vocal (the upper row) and writing (the lower row) activities from
ABC book Lipka [3].
Next, an attention is given to vocal activities. A typical scenario is a short text
which must be read by children aloud. This can be handled same manner as an reading
tutor [8]. Children read word by word in microphone and a speech recognizer evaluate
accuracy of them. It is not only due to speech recognition accuracy, which is better for
isolated words, but the 1st graders learn to read whole isolated words at first. A typical
recognition vocabulary includes all words from the story which size is not very big (cca
200) - tight vocabulary. The recognizer uses HTK toolkit [14] and achieves quite good
speaker independent word accuracy [8]. Application time responses are less than 1 sec.
A recorded isolated word is cut and recognized as a whole. Children can also listen to
text which is prerecorded. The recording is forced aligned to text so that children can
see an actual spoken highlighted word of the text [9, 10, 13].
It seems that an activity where children entitle pictures can be handled same way.
But a problem arose. While the reading tutor can use limited tight recognition vocabulary the entitle activity has to use a big one. To cover all Slovak words with their forms
the vocabulary must be sized at hundreds of thousands. An ordinary expansion of the
HTK recognizer leads to problems with time responses. To avoid this problem the main
vocabulary can be reduced by an approximation of recognized phoneme sequence at
the beginning. It will be good if the reduced vocabulary contains a right word. It will
give chance the recognizer to choose it.
257
258
Marek Nagy
SAMPA Slovak word SAMPA
Vowels
a
papier
a:
E
pero
E:
I
pivo
I:
O
popol
O:
U
puto
U:
{
mäso
Diphthongs
I_ˆa piatok
I_ˆE
I_ˆU/ cudziu
U_ˆO
Consonants
p
popol
b
t
vata
d
c
platit’
J/
k
páka
g
ts
maco
dz
tS
mačatá
dZ
f
fajka
v
f_v vdova
U_ˆ
s
osem
z
S
košel’a
Z
x
chata
h/
G
vrch hory
j
dvaja
I_ˆ
r
para
r=
r=: vŕba
l
skala
l=
l=: vĺča
L
m
mama
F
n
vrana
n
N
cengat’
J
Slovak word
pás
nové
pískat’
pól
púpava
spievat’
kôň
žaba
voda
hád’a
agát
hádzat’
hádžem
slovo
pravda
váza
veža
noha
kraj
prst
vlk
l’avý
amfiteáter
inžinier
vaňa
Table 1. Slovak SAMPA symbols with examples
2 Phoneme sequence recognition
Hidden Markov Models (HMM) of Slovak monophones and triphones have been trained
during speech recognition training procedure [8]. Parameters of used data and the recognizer accuracy are depicted in Table 2. The baseline recognizer uses tight vocabulary
and triphone (context phone) HMM models.
The data are from a project Multimedia reading book [13]. Speakers are children
at age 6-12 and patterns were taken in different uncontrolled conditions (different microphones, high SNR, . . . ). Used phonemes are showed in Table 1 with word examples
and were chosen according to Slovak SAMPA [4].
The phonemes recognizer uses one loop grammar of all Slovak phonemes. To improve accuracy a bi-phone (bi-gram) matrix has been computed from training data.
Using Speech and Handwriting Recognition in Electronic School Worksheets
data speakers wavs words accuracy%
train
352 35934 8293
92.1
test
23 2705 474
86.4
Table 2. The baseline speech recognition accuracy.
Accuracy %
phoneme models
triphone models
data
with bi-gram
with bi-gram
train 76.9 (80.5) 78.5 (81.8) 95.2 (96.0) 95.6 (96.2)
test 79.8 (84.3) 80.4 (84.9) 88.3 (90.5) 88.1 (90.5)
Table 3. The accuracy of individual phonemes. Numbers in parenthesis show an correction after
introduction long and short vowels equivalency.
Accuracy %
phoneme models
triphone models
data
with bi-gram
with bi-gram
train 12.8 (15.6) 22.0 (26.7) 52.0 (53.4) 65.1 (66.8)
test 4.7 (6.6) 12.2 (16.9) 11.3 (11.9) 21.0 (22.4)
Table 4. The accuracy of whole sequence of phonemes. Numbers in parenthesis show an correction after introduction long and short vowels equivalency.
When the typical triphone training procedure is retraining isolated phoneme models
the forced align algorithm is used to determine time boundary of phonemes [14]. The
best phoneme sequence is chosen from several ones in phonetics dictionary (HVite
command [14]). So transcribed training data are used to compute “bi-gram” statistics
matrix. The HLStats command from HTK [14] is used. Then the matrix is included into
the recognition “loop” network.
3 Phonetic dictionary searching
The recognized phoneme sequence is converted into word candidates in the next step.
Whole phonetic dictionary [5] of 82180 words is searched to choose candidates. The
Levenshtein distance algorithm is used for this purpose. But possible operations (insertion, deletion, substitution) are weighted.
3.1 Insertions and deletions
The insertions and deletions are basic operations. It is important which phoneme is actually being deleted. It can happen that vowels are duplicated in the phoneme sequence.
It is not so significant for similarity and therefore weight of this deletion or insertion
is defined as 1. A special situation is with composite phonemes dz, dZ, ts, tS. The
consonants d, t last short time and the phoneme recognizer can only detect the second
259
260
Marek Nagy
part. It is depicted in Table 5 as a similarity (dz ↔ d and dz ↔ z and so on) because it
is made as a special substitution rule in an application.
dz ↔ d, z
dZ ↔ d, Z
ts ↔ t, s
tS ↔ t, S
Table 5. Insertion and deletion in composite phonemes. (weight 1)
All vowels, diphthongs and unstopped l=, r= are inserted or deleted with weight 3
because they are base elements of a word (syllable). Other cases have weight 2.
3.2 Substitutions
Sometimes the phoneme recognizer has problems with vowel duration. It has problems
to correctly determine which variant of a vowel is better: short or long. And therefore
it will be good to consider both variants. Rules are given in Table 6. The starting experiment also shows that accuracy is increasing. See Table 3 and the accuracy in the
parenthesis. The duration can not be ignored. For example Slovak words zastávka ↔
zástavka differ only in the vowel duration. Besides vowels there are unstopped consonants r, l too.
a ↔ a:
E, ({) ↔ E:
I ↔ I:
O ↔ O:
U ↔ U:
l, (l=) ↔ l=:
r, (r=) ↔ r=:
Table 6. Substitution variants based on duration. (weight 1)
A special case is phoneme { which is not used in ordinary Slovak language now
and therefore it can be substituted by the phoneme E. Due to too little training patterns
the phoneme G must be replaced by phoneme x. The phoneme l= is same case. The
replacement rules are in Table 7.
According to the book [6] the first two formats of vowels can be plotted such as in
Figure 3. A tongue position can approximately be plotted same way as the formants. If
a tongue do not take the exact position then the recognizer can make mistakes between
neighboring vowels. This aspect is covered by Table 8. The weight of these shifts is 3.
A typical phonemes exchange is based on pair consonants that differ only by presence of voice [5]. See Table 9.
Using Speech and Handwriting Recognition in Electronic School Worksheets
{↔E
G↔x
l= ↔ l
Table 7. Ignored variants - replaced phonemes (weight 0)
Fig. 3. Vowels continuity diagram (solid line - according to formants, dotted inner line - according
to tongue position)
I↔E
E↔a
a↔O
O↔U
Table 8. Substitution variants based on vowel shifts (weight 3)
b↔p
d↔t
J/ ↔ c
g↔k
f_v ↔ f
z↔s
Z↔S
dz ↔ ts
dZ ↔ tS
(G), h/ ↔ x
Table 9. Substitution variants based on pair consonants (voiced/unvoiced) (weight 1)
Table 10 is showing rules based on an assimilation aspect of Slovak language. The
assimilation can have a cross words effect. A typical example is u-v assimilation (e.g.
strojov). The f-v (e.g. včela) assimilation has own alophone f_v but speakers tend to
speak it either as f or as v. Other type is i-j assimilation in diphthongs especially (e.g.
piatok, ahoj). The next n-N assimilation rule is based on an inadequate pronouncing.
And the last rules are based on non-literary pronouncing (e.g. môj, mój, moj).
261
262
Marek Nagy
U ↔ U_ˆ ↔ v ↔ f_v
v↔f
I ↔ I_ˆ ↔ j
n↔N
m↔F
U_ˆO ↔ O:, O
Table 10. Substitution variants based on assimilation (weight 1)
d ↔ J/
t↔c
n↔J
l↔L
z↔Z
s↔S
dz ↔ dZ
ts ↔ tS
Table 11. Substitution variants based on palatalization (weight 1)
The palatalization is other aspect of Slovak language. It is marked by a hiccup above
consonant graphemes or not if graphemes e, i follow after d, t, n, l. This hiccup omitance
makes problems because it can be read hard or floppily. Also some Slovak dialects
have tendency read it non-literary hard everywhere and some speakers read it floppily
everywhere. The palatalized pairs are collected in Table 11.
A special case is when a composite phoneme is demerged into two parts. It requires
to include a new operation in Levenshtein algorithm. The substitution step can go
through two symbols but only in one direction (horizontally or vertically). See Figure
4. All cases are depicted in Table 12.
Fig. 4. Example of the new operation (decomposition) of composite phonemes. The arrow can go
through two cells either horizontally or vertically.
Two substituted vowels or diphthongs which are not among mentioned rules will be
substituted with weight 5. A vowel ↔ diphthong substitution has weight 7. All other
not mentioned substitutions have default weight 3.
Using Speech and Handwriting Recognition in Electronic School Worksheets
I_ˆa ↔
I, I_ˆ, j + a
I_ˆE ↔
I, I_ˆ, j + E
I_ˆU/ ↔
I, I_ˆ, j + U
U_ˆO ↔ U, U_ˆ, v, f_v + O
dz ↔
t, d + dz, z
dZ ↔
t, d + dZ, Z
ts ↔
t, d + ts, s
tS ↔
t, d + tS, S
N↔
n + g, k
Table 12. Substitution variants based on decomposition of phonemes (weight 1)
4 Results
The new isolated word recognizer is assembled from three stages. At the first, the phonetic sequence is estimated. At the second, the N-best candidates are searched in the
phonetic dictionary [12]. And at the third, one best candidate is chosen. The accuracy
of the recognition and a time per a word are showed in Table 13. The reached accuracy
of the three stage recognizer is comparable with a typical HTK recognizer which uses
a beam width search heuristic [14]. Of course it was a goal that the accuracy will not
be worse. But time responses are a little better. Three summed numbers in Table 13
represent time consumption of the individual stages. The time reduction achieves 36%.
Some time reserves can be seen at the phonetic dictionary searching stage. Now the
complete dictionary is saved on a remote MySQL database server. This database server
is used by other applications too. The search time can be reduced by rational local storing in computer memory. A little (not mentioned) experiments show reductions close
to 1 second. In practical application it will be better connect all three stages together
closer and reduce response time to minimum 1 second.
Only two variants of N-best candidates were chosen (50-best, 100-best) for comparison. The 100 candidates are sufficient as can be seen in Table 13. It seems that
fixed number of candidates is not so good. Now two words with same distance can
be separated one among the candidates and one outside of them. In future it will be
better to use distance criteria too. So that all words which have same distance will be
altogether among candidates or not.
through 50-best through 100-best
basic
basic + beam w.
data words Accuracy sec/w Accuracy sec/w Accuracy sec/w Accuracy sec/w
test 82180
66.4 3+5+1
67.2 3+5+1
66.1
63
67.6
14
Table 13. The accuracy of the three stage recognition. Triphone models with bi-phone statistics were used. Also the basic recognizer with the beam width heuristic [14] is presented for a
comparison.
263
264
Marek Nagy
5 Conclusions
In this paper the problems of electronic school worksheets are analyzed. The goal is to
transfer paper worksheet exercises, which are developed by pedagogic professionals,
from paper into computers and enrich them by additional functionality. The ABC book
Lipka [3] was taken as a basis where exercises are more allowable to this idea. Two
types of activities were identified: vocal activities and writing activities. In this article
the typical entitle exercises were enriched by the speech recognition approach. The
electronic worksheet can check if kids speak correct words that represent the pictures.
The isolated word speech recognition is used for this reason. Because kids can speak
random Slovak words the recognizer’s dictionary must be extended. In this article the
whole phonetic dictionary contains 82180 words that are in base forms mostly. The
Slovak paper phonetic dictionary has been adopted [12]. This caused increasing the
time response. Therefore the better search approach is presented in this contribution. It
reduces the time response by more massive vocabulary reduction than the typical Viterbi
beam width search approach [14]. The obtained time reduction is significant and with
better phonetic dictionary setup it can be close to 1 second per a word in future.
Acknowledgments
This work was partially supported by the Slovak Scientific Grant Agency (VEGA)
under the contract No. 1/4060/07 and by the Slovak Research and Development Agency
under the contract No. LPP-0056-07.
References
[1] Červeň, J.: Writing tutor – computer helps children learn to write, In Proceedings of the 9th international conference on Informatics, Informatics 2007, SSAKI,
Bratislava
[2] Hirková, A., Nemčíková, M.: Pracovný zošit pre 1. ročník základných škôl k
šlabikáru Lipka, Aitec, Bratislava 2008
[3] Hirková, A., Nemčíková, M.: Šlabikár Lipka pre 1. ročník základných škôl, Aitec,
Bratislava 2008
[4] Ivanecký, J., Nábělková, M.: Fonetická transkripcia SAMPA a slovenčina,
Jazykovedný časopis, Vol. 53, No. 2, 2002, Bratislava, pp. 81 – 95
[5] Král’, Á.: Pravidlá slovenskej výslovnosti, (tretie vydanie 1996), SPN, Bratislava
1983
[6] Král’, Á, Sabol, J.: Fonetika a Fonológia, (prvé vydanie), SPN, Bratislava 1989
[7] Lipková, J.: Počítač ako učitel’ písania (Computer as an writing tutor), master’s
degree thesis, Comenius University, 2008, Bratislava
[8] Nagy, M.: Čítanie s tučniakom, využitie rečového komunikačného rozhrania pri
výučbe čítania detí na I.stupni ZŠ, In Proceedings of the 7th international conference, APLIMAT 2008, Bratislava, Slovakia, pp. 101 – 110
[9] Nagy, M.: Multimediálna čítanka 2008, 4.ročník, Active Advice, Bratislava, 2008,
273 str., ISSN 1337 5601
Using Speech and Handwriting Recognition in Electronic School Worksheets
[10] Nagy, M.: Počítač dokáže naučit’ plynulo čítat’, Učitel’ské noviny, č. 2, 2008,
Bratislava, str. 31
[11] Tomacha, M.: Počítač – učitel’ písania (Computer – a writing tutor), master’s
degree thesis, Comenius University, 2007, Bratislava
[12] Vančo, P., Nagy, M.: Creating of Slovak Electronic Phonetic Dictionary for Use in
Speech Recognition, In Proceedings of the 3rd international conference, Slovko
2005, Bratislava, pp. 216 – 219
[13] (www): Projekt Multimediálna čítanka,
http://cpr.ii.fmph.uniba.sk/citanka
[14] Young, S., Evermann, G., Kershaw, D., Moore, G., Odell, J., Ollason, D., Povey,
D., Valtchev, V., Woodland, P.: The HTK Book Version 3.2, Cambridge, England,
Cambridge University, 2002
265
Composite Lexical Units as an Element
of Lexicographical Historical Computer System*
Irina Nekipelova
Izhevsk State Technical University, Russia
[email protected]
Abstract. This article is devoted to the development in the field of modelling
the description of a word lexical value in the information-search system "Manuscript". This aspect is connected with a problem of use a system for realization
of linguistic researches in the field of lexics and semantics and working out the
linguistic search system allowing the user to have exact idea of a word lexical
value and its semantic connections in language and texts of ancient
manuscripts, stored in a database.
Keywords: history of Russian language, grammatical semantics, triodion.
1 Introduction
One of the way to save and accumulation information is presentation and research of
ancient script monument in special portals. This is a multiple-aspect appearance, an
important factor in humanization of social relation and in process of forming of cultural and scientific picture of the world.
Creation of full-text database, forming of corpora of texts Russian early manuscripts and their researching favour the decision of some of important tasks for
linguistic science: 1) electronic publishing of ancient Slavonic script monuments
favoures the saving of some part of cultural and scientific heritage that keeps in the
vaults of libraries in poor condition that’s why they don’t available for researching
often; 2) lingo-textual description of different time ancient lists is a necessary element
for study of ancient script monuments as important part of national culture component; 3) different aspect research of lexical composition and semantic relations of
language of ancient Russian texts favoures more deep and detailed understanding of
history of language and history of the Slavs life; 4) studies in lexical and grammatical
semantics favoures the research of peculiarity of functioning language units in ancient
texts in different period of language development because language semantics is an
important element in forming of the Slavs linguistic picture of the world.
*
Работа выполнена в рамках аналитической ведомственной целевой программы
Федерального агентства по образованию Российской Федерации “Развитие научного
потенциала высшей школы (2009-2010 годы), проект “Лингвотекстологические и
корпусные исследования грамматической семантики древнерусского текста”
(регистрационный номер 2.1.3/2987).
Composite Lexical Units as an Element of Lexicographical Historical Computer System 267
Electronic publishing of transcription of Slavonic triodion early manuscripts and
their lexico-semantic and grammar-semantics peculiarity research presuppose to
development and adoption a landmark method of account of vocabulary with a glance
of its system and functional properties and words semantic relations. Results of
research will have an applied significance for lexicography, lexicology, semantics and
history of language because system of the Old Russian language lexical units which
use in modern lists of triodions is that of as an adjuvant to Russian language teaching
practice and attracting attention of scientists and all comers to that little-investigated
manuscripts.
The most promising technique of manuscripts integrated study is electronic collection creating to full-text database. Quality of that collection is in saving information
content, fast-access retrieval, full-featured of date sample and sort, possibility of
replenishment a database with some new information and visibility of date in the
Internet.
Electronic publishing of transcription of Slavonic triodion early manuscripts and
their lexico-semantic and grammar-semantics peculiarity research presuppose to
development and adoption a landmark method of account of vocabulary with a glance
of its system and functional properties and words semantic relations. Results of
research will have an applied significance for lexicography, lexicology, semantics and
history of language because system of the Old Russian language lexical units which
use in modern lists of triodions is that of as an adjuvant to Russian language teaching
practice and attracting attention of scientists and all comers to that little-investigated
manuscripts. Thereat in information analysis system “Manuscript” (http://manuscripts.ru/) create databases which include collections of different genres the Slavic script
monuments. Collection of triodion creates in their midst.
Nowadays collection includes such publication: Festal Triodion, notificationly, the
end of XII century (ГИМ, Воскр. 27 перг.), 206 p.; Lenten Triodion, notificationly,
XII century(ГИМ, Син. 319), 315p.; Festal Triodion, services of Twelve Great Feasts
and for chosen saints, the beginning of XII century, the end of XII – the beginning of
XIII centuries and XIV century(РГАДА, ф. 281 (Син. тип.), № 139), 208p.; Festal
Triodion, XIIcentury, РГАДА, ф. 381, № 138),173 p.; Lenten Triodion, the end of
XIII – the beginning of XIV centuries (РНБ, Погод. 41), 157 p. Expected expansion
of triodions collections within the bounds of intended to electronic publishing such
manuscripts: Lenten and Festal Triodion (Triodion by Moses Kiyanin), the end of XII
– the beginning of XIII centuries (РГАДА, ф. 381, № 137), 257 p.; Feasts with insertion from Festal Triodion, the second part of XVIII century (РНБ, Соф. 385), 150 p.
About importance the text of Triodion by Moses Kiyanin for study of Russian language history wrote A.I. Sobolevskii [1]. Most scientists thought that this manuscript
one of the most significant creation in the period of XI-XIV centuries.
Texts of manuscripts must be presented integratly showing general text information in the movement of Slavic culture. So-called virtual libraries give an opportunity
268
Irina Nekipelova
not only to view documents, but receive information about their structure, state,
description, etc.
Use of this system makes it possible to get material at short notice for linguistic,
historical, philological and lingo-textual research, its full and exact lexico-semantic
analysis that give an opportunity a wide range of users IAS “Manuscript” to go beyond the scope of the research of single text and/or contemporary texts and to draw a
significant fundamental conclusion on the ground of text collections or corpus of Old
and Middle Russian periods.
Electronic publication is necessary to direct usage of gathered scientific matter for
linguistic problems solution in the range of lexicon and semantics of the Old Russian
language and for linguistic research of triodions lists, i.e. for analysis and modelling
lexical matter, which meets in triodion. So, within the bounds of linguistic research of
used vocabulary analysis is carried out within the range every triodion and description
lexical and semantic the Old Russian word relations that presented in every list of triodion. Lexical description and electronic historical dictionary creation are necessary
first of all for modern thinking more understandable of language of manuscript’s text
and come across for users of reading. Thus for development of multifunctional webmodules that contained publication of ancient and medieval Slavic manuscripts;
developments in the range of lexical meaning of the word and its bounds in the range
of semantics modelling are actual.
2 Semantic model
Types of lexical description are foundation of lexical meaning modelling and word’s
semantics in database. It should be noted that all meaning types, considered as complementary for each other, i.e. as a constituent (side, aspect, part of all). It’s important
because some facts, that complicated research, must be considered in describing
word’s semantics in its history. First of all, word has remained to this day only in the
certain movement that hampers fixing all possible usages of word and its bounds with
other language units, so long as extant texts cannot reflect all word relations, that realized in any period of language’s development. Secondly, in interpreting of bounds of
words that use in manuscripts cannot always be used to say about that analysis absolute adequacy, i.e. description of word’s semantics is regarded through the alembic of
availability of word’s semantic bounds in modern language which might not be at that
lexeme at earlier period of language development.
Nowadays, structure of word’s description was developed that was presented in
the form of hierarchical bounds of words. Modelling of semantics is development of
typical structure of semantic description, which structure’s usage of capacity is made
conditional on individual characteristics and relations of words.
At first, nominative type of sense is worked out by us, because this type represents
lexico-semantic bounds of words. Nominative sense is presented with meanings of
non-structure lexical units: lexemes and phraseological units. In this case we tell
Composite Lexical Units as an Element of Lexicographical Historical Computer System 269
about lexical and phraseological sense, respectively. For lexical description of word
leading is identification of lexical meaning of word, in which seven basic types of
description of lexical meaning of word are presented such as encyclopaedic, explanatory, etymological, synonymous, antonymous and homonymous. For phraseological
description is important phraseologation of component units in history of language,
i.e. their relation to vocal formulas which are spreaded in early stage of language history, or to phraseological units proper, which are appeared at a later period.
Dictionary must contain references to all usage of concerned lexeme in context,
i.e. full information about all usage in any context.
Let’s present this structure of description by example of word water, which was
reiterated made of in text of Festal Triodion, XI-XII century (РГАДА, ф. 381 (Син.
тип.)), №138.
Nominative sense:
1. lexical sense:
1.1. defining sense:
1.1.1. linguistic sense: вода, ж. "вода, естественная влага" [2]. К этому разделу
словаря должны быть отсылки всех употреблений слова вода в следующих
контекстах:      (114 л.), 
 (112 л.),        
  (039 л.),     (052 л.), 
         (023
об.).
1.1.2. context sense: бессмертные воды, вода бессмертия – то же, что "живая
вода, животная вода":   (112 л.),   






 (116 л.); вода премудрости:   
       
      (116 об.),
       
(114 л.),    (114 л.); воды веры –
то же, что "источник веры":       
         
        
(114 л.); воды спасения – то же, что "воды веры":    
           
       
  (114 л.); неповинная вода – то же что "чистая вода,
нежертвенная":      
          
   (041 л.);
1.2. encyclopaedic sense: вода (оксид водорода), простейшее устойчивое
химическое соединение водорода с кислородом, Н2О <...> [3].
270
Irina Nekipelova
1.3. etymological sense: 1) вода, -ы, ж. - «бесцветная, более или менее
прозрачная жидкость, являющаяся главной составной частью гидросферы
земного шара, образующая реки, озера, моря, океаны». Прил. водный, -ая, -ое,
водяной, -ая, -ое, отсюда водянистый, -ая, -ое. Укр. водá, вóдний, -а, -е, водянúй,
-а, -е, водянúстий, -а, -е, вóдявий, -а, -е; блр. вадá, вóдны, -ая, -ае, вадзяны, -áя,
-óе, вадзянíсты, -ая, -ае; болг. водá, вóден, -дна, -дно, воднúст, -а, -о, воднúкав, -а,
-о; с.-хорв. вòда - «вода», «река», вöднū, -ā, -ō- «водный», «водяной», вöден(ū), -а,
-о, - «водяной», «водянистый», вòдньикав(ū) -а, -о — «водянистый»; словен. voda,
voden : vodni, -а, -о; чеш. voda, vodáсký, -á, -é - «водный», vodnatý, -á, -é –
«многоводный», «водянистый»; словац. voda, vodný, -á, -é - «водный», «водяной»,
vodnatý - «водянистый»; польск. woda, wodny, -а, -е - «водный», «водяной», wodnisty, -а, -е; в.-луж. woda, wódny, -а, -е - «водный», «водяной», wodnawy, -а, -е «водянистый»; н.-луж. wóda, wódny, -a,-е - «водный», «водяной», wodniaty,
wódowaty, -a, -e - «водянистый». Др.-рус. (с XI в.) вода, водьнъ, водьный,
значительно позже (с XVI в.) водяный (Срезневский, I, 276, 279; Доп. , 36). Ст-сл.
вода, водьнъ, водьныи, водьскъ, -а, -о (SJS. 1:5, 205, 207). ▫ О.-с. *voda. И.-е. база
*aued- : *ǔd-; им. ед. *ụédor : *ụódō(r) |> о.-с. *voda]; ср. локатив ед. *udén(i),
род. ед. udnés (см. Рокоrnу, I, 78). На славянской почве родственные
образовании: о.-с. *vydra (см. выдра), о.-с. *vĕdro (см. ведро). Ср. гот. watō «вода»; то же др.-в.-нем. waззar (совр. нем. Wasser); англосакс. wœter (совр. англ.
water); греч. ϋδωρ, фригийск. βεδυ <*uedō); др.-инд. udάn(i), локатив ед., udnάh;,
род. ед. от udakά-m - «вода»; хетт. wātar, род. wetenas; с назализованным
вокализмом: лит. vanduõ, род. vandeñs (при жем. unduõ), латыш. ūdens - «вода»;
др.-прус, unds - тж.; латин. unda - «волна» [4]; 2) водά, сюда же вόдка, укр., блр.
водά, др.-русск., ст.-слав. вода ϋδωρ (Супр.), болг. водά, сербохорв. вòда, словен.
vóda, чеш. voda, слвц. voda, польск. woda, в.-луж., н.-луж. woda. Древние ступени
чередования представлены в ведрó, выдра. || Родственно лит. vanduõ, род. п. vandeñs, жем. unduo, д.-в.-н. waззar «вода», гот. watō, греч. ϋδωρ, ϋδατος, арм. get
«река», фриг. βέδυ, др.-инд. udakάm, uda-, udάn- «вода», unάtti «бить ключом»,
«орошать», ōdman- ср. р. «поток», алб. uj «вода»; носовой согласный в лат. unda
«волна» и лит. vanduõ вторичного происхождения; см. Вальде 850; И. Шмидт,
Pluralb. 202 и сл.; М.-Э. 4, 404 и сл.; Хюбшман 434; Уленбек, Aind. Wb. 28 и сл.
Древняя основа на r/n [5]; Вода. Общеслав. индоевр. характера. Та же основа, но
с перегласовкой, содержится в словах ведро, выдра [6]. Etymological sense is presented in compliance with information from etymological dictionaries.
1.4. homonymous sense: вода – "водное пространство (о море, реке, озере и т.
п.)":        
        
    (049 л.),    
   ( 111 об.),    
         (037
л.).
Composite Lexical Units as an Element of Lexicographical Historical Computer System 271
1.5. synonymous sense: вода (в значении "море") = пучина:  
       (112 л.).
1.6. antonymous sense: вода (в значении "море") – суша, земля: 
        
 (041 л.),     (035 об.).
2. phraseological sense:
2.1. vocal formulas: животная вода:    
      
          
    (115 об.),    
.      (114 л.),
      (114 об.), 
        
         
         (043 об.); вода
бессмертия:  (112 л.); бессмертные воды:  
     
 (116 л.); живая вода:     
(112 л.),    (112 л.); жизненная вода:  
      (115 л.).
2.2. phraseological units which aren’t presented in text of triodion
Let’s suppose that user is interested in meaning of word вода and its lexico-semantics
bounds with other words in the following movement:   
       (039 л.). In that case lexeme
is referred us to part of dictionary, where information about using of that lexeme is
contained. In this case, information presents like that:
Nominative sense:
1. lexical sense:
1.1. defining sense:
1.1.1. linguistic sense: вода, ж. "вода, естественная влага":   
       (039 л.). 
1.1.2. context sense isn’t presented.
1.2. encyclopaedic sense: вода (оксид водорода), простейшее устойчивое
химическое соединение водорода с кислородом, Н2О <...>.
1.3. etymological sense: 1) вода, -ы, ж. – «бесцветная, более или менее
прозрачная жидкость, являющаяся главной составной частью гидросферы
земного шара, образующая реки, озера, моря, океаны» <...>; 2) водά, сюда же
вόдка, укр., блр. водά, др.-русск., ст.-слав. вода ϋδωρ <...>; 3) вода. Общеслав.
индоевр. характера <...>.
1.4. synonymous sense isn’t presented
1.5. antonymous sense isn’t presented
272
Irina Nekipelova
1.6. homonymous sense: вода – "водное пространство (о море, реке, озере и т.
п.)":        
        
    (049 л.),    
   (111 об.),    
         (037
л.).
2. phraseological sense:
2.1. vocal formulas isn’t presented
2.2. phraseological units isn’t presented
However, user may to be interested in meaning of lexeme вода (water), which
homonymous by water word in meaning “limpid liquid” and its lexico-semantics
bounds with other words in the following movement:    
      
         
(049 л.). In this case, user receives the following information:
Nominative sense:
1. lexical sense:
1.1. defining sense:
1.1.1. linguistic sense: вода – "водное пространство (о море, реке, озере и т.
п.)":        
        
    (049 л.)
1.1.2. context sense isn’t presented
1.2. encyclopaedic sense isn’t presented
1.3. etymological sense isn’t presented
1.4. homonymous sense: вода - "вода, естественная влага":  
   (114 л.),   (112 л.), 
         (039
л.),    (052 л.),   
       (023 об.).
1.5. synonymous sense: вода = пучина:    
     (112 л.).
1.6. antonymous sense: вода (= море) – суша, земля:  
       
 (041 л),     (035 об.).
2. phraseological sense:
2.1. vocal formulas isn’t presented
2.2. phraseological units isn’t presented
Such placement of lexico-semantics information about word is required certain
comments.
Composite Lexical Units as an Element of Lexicographical Historical Computer System 273
Encyclopaedic and etymological interpretations are reflected original relations of
word. Encyclopaedic sense is given for reference value of this word. All homonymous lexemes haven’t encyclopaedic sense because this type of dictionaries hasn’t
senses of homonym. The same situation with etymological description of word –
derivative homonyms haven’t etymological characteristic. So, we see in diverse terms
of descriptions and senses of word which don’t contradict to each other.
Explanatory sense is presented in the form of the linguistic and context sense. Linguistic sense of word is sense, as a rule, that is fixed in Dictionary of Russian
language of XII-XVII centuries and agreed with etymological sense. Senses, that isn’t
fixed in dictionaries, is described as context senses of significance lexical units with
semantic characteristics which they have in reduced context directly.
It’s important under interpretation of word to identification of context senses of
word. Basic criteria of differentiation of linguistic and context senses of word, consecutive usage of them to description of word functioning reveals linguistic or context
character of word meaning, are worked out by us.
Special role in development of word’s semantics plays its relations with other
words, including stability usage of the word in the define movement.
Semantics of the Old Russian language researching presupposes researching of the
important component – grammatical semantics. Grammatical semantics is the integral
component of semantics in whole which enables to supplement integral linguistic picture of the world that is reflected in texts of Old and Middle Russian periods.
Grammatical semantics must be seen as separate bloc of textual usage of language
unit’s researches in ancient scripts monuments that binds semantics and grammar in
whole. It is very important under functioning of composite language units in
manuscripts.
Grammatical semantics of vocal formulas and phraseological units are attracted
particular interest because the define role in development of word’s semantics plays
relations with other words including stability unstability (lexical and grammatical)
usage in the movement, degree of unification of components formation.
Grammatical semantics of set expressions regards of forming of meaning and
composition of set expressions and phraseological units. Generally in that period scientifics fix functioning of vocal formulas. Compound lexical units are special
lexicographical units that have special characteristics. In whole vocal formulas permit
to transposition, to substitution, to admission of components with saving of initial
sense, even so grammatical variativity used of components is met often, which may
lead to change of semantics.
Description of phraseological sense of word is hard task; because many factors
must be take into account such as time of appearance of the formation, field of usage
and conversion degree, degree of activity and degree of repeatability in works of different genres, expansion narrowing of phraseological sense and degree of unification
of the forming components. It task becomes more complex with respect to the Old
Russian language. As is generally known, process of phraseologization has long his-
274
Irina Nekipelova
tory and that set expressions, which are met in texts of scripts monuments of XI-XIV
centuries, aren’t phraseological units. Generally, in that period scientists don’t fix of
functioning of vocal formulas.
V. Deryagin reasonable notes: “Для периода сложившейся деловой письменности в языковом отношении под формулой целесообразно понимать
фразеологизм номинативного или коммуникативного характера, а также
словосочетание, синтаксическую конструкцию (модель предложения) с более или
менее постоянным лексическим составом. В отдельных случаях формула может
состоять из нескольких предложений, связанных между собой синтаксически и по
смыслу” [7]. Formula is a basic unit of stylistic analysis of official text. It is unit of
text level and with it formula may be define in terms that use for units in other levels
that lower in hierarchy: formula is sentence (definite type), formula is set expression
(definite type), a phraseological unit [8].
Well, it should be affirmed that availability of linguistic formulas not only in official written language texts, but in text of different genres, because vocal formulas
using is defined not only with characteristics of genre, but with general linguistic processes. One of the means of forming of vocal formulas is semantic tracing of Greek
metaphors, which is led to its symbolization (look in formings ,
and so forth). Term formula V.V. Kolesov correlates with wording of a borrowed symbols of Greek culture in Old Russian texts. “Древнейшие
[заимствования] не были свободны от контекстов, в составе которых они
перешли к славянам, и эти контексты попадали к ним в письменной форме
переведенных текстов. Заимствовалось словосочетание целиком, почему и
заимствованные слова оказывались фразеологически связанными” [9].
In the capacity of working-class definition in our research we take which suggests
N.A. Antadze. Formula is a special set expression; scientist attributes them to phraseological combinations by common feature of invariability common meaning and
relating usage in examples. Formulas admit of shift, substitution, passing of components with preservation of invariable common meaning [10]. So, often formulas are set
expressions that are bound syntactically (in the context) and phraseologically (by
sense, content), seldom formulas are sentences which are characterized by stability
and repeatability [11].
Structure of detection of phraseological possibility of word demands some
comments. In semantic model “vocal formulas” field always fills. It should be noted
that lexical units which are marked in that field not always have fixed in different dictionary sense; in field of “phraseological units” is empty in terms of usage of capacity
by citations from text of triodion, but in that field dates are showed in phraseological
dictionary, which reflects well-established phraseology of present language. It is natural that this two fields do not agree and haven’t common date.
Fields, where phraseological sense is fixed, fill as far as possible bulking of
research materials.
Composite Lexical Units as an Element of Lexicographical Historical Computer System 275
In the capacity of material of research of grammatical semantics wordforms are
taken, such elements of composite units that met in text of Festal Triodion XII century
(РГАДА, ф. 381, №138, 173 л.). Compound units of language with stable lexical-syntactic character with a component  had been revealed in the Festal Triodion:
Животная вода in the meaning of “well, running water, drinkable water”
[Словарь 1975, 2: 249]:      .
     (114 л.);   
   (114 об.);    
      
          
    (115 об.);    
        
         
       (043 об.);
Живая вода in the meaning of “well, running water, drinkable water” [Словарь
1975, 2: 249]:      (112 л.);  
 (112 л.);
Жизненная вода the same as живая вода/животная вода in the meaning of
“drinkable water” [Словарь 1975, 2: 249]:   
     (115 л.);
Бессмертные воды the same as the живая вода/животная вода in the meaning of
“well, running water, drinkable water”:    
     (116 л.);
Вода бессмертия in the meaning of “drinkable water”:  (112
л.).
In whole we say only about two vocal formulas which have lexico-grammar variations:
1. We say about functioning of vocal formula живая вода/животная
вода/жизненная вода in the meaning of “well, running water, drinkable water”
which composed of only the first component is changed: живая, животная,
жизненная. Meaning of that component goes back to meaning of word жизнь. It
assigns meaning of all compound units – “water that needs for life support”.
«Единство составного имени основывалось на единстве обозначаемого им
понятия. Грамматически подчиненным компонентом являлось прилагательное, а
в качестве семантически подчиненного компонента выступало существительное,
поскольку оно определяло лишь частичную принадлежность образования...
Прилагательные несли основную семантическую нагрузку» [12]. It should be noted
that substitution of vocal formula’s component doesn’t change its meaning. All formula’s lexical variations are used to in similar and identical contest. It should be said that
variation жизненная вода is used more infrequent.
Grammatical variation, that is observed in functioning of this compound units of
language in the text, is expressed: 1) reverse order of components in composed of for-
276
Irina Nekipelova
mula, for example, вода животная:     
.      (114 л.);
      (114 об.) preorder животная вода:
          
   (115 об.); 2) using of adjective forms that don’t segment instead
of segment forms, for example, вода животна/воды животны:  
       
           
  (043 об.), well as вода жива:    
 (112 л.);    (112 л.) and вода жизньна: 
       (115 л.).
2. The next vocal formula, that we find in the text of Festal Triodion is
бессмертные воды. Meaning of that language unit didn’t fix in historical dictionary.
This meaning associated with definite of drinking water, but it isn’t become formed
with word жизнь, but with смерть. In semantic point the last formula is oppose to
the formula живая/жизненная/животная вода. If the second formula has object
meaning “drinking water”, then the first formula бессмертная вода has in the text a
symbolic significance, i.e. “not only for life supporting, but for deliverance from
death, acquisition of eternal life”. That semantic difference with difference of lexical
components, i.e. we say about two different vocal formulas, not only one. Lexico-syntactic meaning will be more and more parted when be used often.
Grammatical variation composed of that formula is presented with syntactic relation’s change of components of compound element. Noun construction immortal
water with agreement bound with main noun and dependent adjective:  
     
  (116 л.);  (112 л.). It substitutes construction вода бессмертия with control bound with main noun and dependent noun:
 (112 л.). It is important that grammatical change compound of
that formula don’t reduce to semantic change in the structure of word meaning.
Analysis of functioning of vocal formulas in texts of old written records allowed
to state a fact of usage incomplete structure. But observable vocal formulas are characterized by only full structure. This is due to the fact that loss compound of leading
in semantic point of adjective leads down to disappearance of the formation, i.e. it is
been unrecognizable in the text.
In whole the process of phraseologization of compound language units is characterized by long history. In many cases vocal formulas were become phraseologies and
had more abstract meanings with the lapse of time. But part of vocal formulas, which
were used in text of XI-XIV centuries, were lost and been absent in nowadays language system. All formation that were researched don’t fixed in phraseological
dictionaries. Only живая вода is fixed and as a result it has meanings: Живая вода
1. Фольк. Мифическая чудодейственная жидкость, возвращающая жизнь
Composite Lexical Units as an Element of Lexicographical Historical Computer System 277
мёртвому телу. 2. Экспрес. Всё, что одухотворяет, благотворно действует,
пробуждает интерес [13].
Modern meanings of phraseology живая вода are bound with reference value.
The object meaning and the subsequent symbolical meaning giving rise to that
phraseologycalized the abstract meaning. It should be noted that lexical and grammatical variations of our phraseology don’t observe nowadays. We see that in
phraseology dictionaries.
Grammatico-semantic analysis of compound language units functioning in old
ancient script texts is important for reconstruction of relations and bounds of words in
the Old Russian. Therewith the analysis is important for modern lexicography system
development, because it shows the permanent development of language system. It
reflects specific character of ancient Slavs language world picture.
3 Conclusion
It should be noted, that all content words have all presented characteristics, signs and
bounds. But only some of that words maximal realize this signs. Analysis of lexical
definitions mustn’t overshadow a real vocal using, contest and realizable lexical units
functions, actualizations of expansion or contraction a conception and forming of context meanings, i.e. somewhat that composes a variable part of language.
Necessary part of research is detection of lexemes that are used in written records.
Lexemes and their meanings didn’t fix in Old Russian dictionaries. It is important not
only for triodions description as independent genre, but for changing of conception
about manuscripts are created in XII-XIII centuries.
In the issue of all word analysis that usage in triodions must be create hierarchical
lexico-semantic language system model; lexicography field that included all bounds
and word relations of Old Russian language too, because totally of meanings are
formed semantic language system, i.e. system of meanings.
References
[1] Соболевский А.И. Труды по истории русского языка. Т.1.: Очерки из
истории русского языка. Лекции по истории русского языка / Предисл. и
коммент. В.Б. Крысько. – М.: Языки славянской культуры, 2004. – С. 12.
[2] Словарь русского языка XI-XVII вв. В 17-ти т. Т.2. – М.: Наука, 1975. – С.
249.
[3] Большая Российская энциклопедия: В 30 т. Т.5. Великий князь –
Восходящий узел орбиты. – М.: Большая Российская энциклопедия, 2006.
– С. 481-483.
[4] Черных П.Я. Историко-этимологический словарь современного русского
языка: В 2 т. / П.Я. Черных. – М.: Рус. яз., 1999. Т.1.: А-Пантомима. – С.
159.
278
Irina Nekipelova
[5] Фасмер М. Этимологический словарь русского языка. Т.1. (А-Д) / М.
Фасмер. – М.: Прогресс, 1986. – С. 330.
[6] Шанский Н.М., Краткий этимологический словарь русского языка / Н.М.
Шанский, В.В. Иванов, Т.В. Шанская. – М.: Просвещение, 1971. – С. 85.
[7] Дерягин В.Я. Варьирование в формулах русской деловой речи XV-XVII
вв. / В.Я. Дерягин // Восточные славяне: Языки. История. Культура: К 85летию академика В.И. Борковского. – М.: Наука, 1985. – С. 243.
[8] Дерягин В.Я. Варьирование в формулах русской деловой речи XV-XVII
вв. / В.Я. Дерягин // Восточные славяне: Языки. История. Культура: К 85летию академика В.И. Борковского. – М.: Наука, 1985. – С. 243.
[9] Колесов В.В. Философия русского слова / В.В. Колесов. – СПб.: ЮНА,
2002. – С. 201.
[10] Антадзе Н.А. Лексико-фразеологический состав Судебников 1497 и 1550
гг.: Автореферат диссертации на соискание ученой степени кандидата
филологических наук / Н.А. Антадзе; Тбил. гос. ун-т. - Тбилиси, 1965. –
С.16–18.
[11] Некипелова И.М. Метонимическая и метафорическая деривация
в истории русского языка (на материале памятников деловой письменности XI-XVII веков). Автореферат диссертации на соискание ученой
степени кандидата филологических наук. – Казань, 2005. – С. 4.
[12] Некипелова И.М. Метонимическая и метафорическая деривация в истории русского языка (на материале памятников деловой письменности XIXVII веков). Автореферат диссертации на соискание ученой степени
кандидата филологических наук. – Казань, 2005. – С.11–12.
[13] Фразеологический словарь русского литературного языка: В 2 т. / Сост.
А.И. Фёдоров. Т.1.: А-М. – М.: Цитадель, 1997. – С. 87–88.
IT: Moving Towards Real Multilingualism
Antoni Oliver and Cristina Borrell
Universitat Oberta de Catalunya, Spain
Abstract. The Linguamón-UOC Chair in Multilingualism of the Universitat
Oberta de Catalunya has developed a project consisting of the automatic elaboration of linguistic resources, those including Catalan. For this, we have
created an automatic extractor of terminology, which is freely distributed,
multi-platform and adaptable to users' needs. One of its most important useful
applications is the elaboration of glossaries, both monolingual and multilingual, based on a set of documents. The system automatically extracts the term
candidates from the texts introduced by the user. Later, a specialist reviews
the list of lexical units to verify the result. In the Chair in Multilingualism we
have applied this tool to the expansion of the Eurovoc glossary with its
Catalan version. We have extracted the equivalents for this language from the
entries in Spanish and a bilingual Spanish – Catalan parallel corpus. From the
application of the extractor, we obtained a multilingual glossary of 2 531
terms involving EU policies. In a second stage, we have used different techniques. Firstly, starting from the terms that were not yet translated by the first
step, we have generated hypothesis of the Catalan translation of the Spanish
Eurovoc terms, thus by the compositionality method. Secondly, we have
applied statistically automatic translation, and finally we have checked the
results in a monolingual corpus or by means of a restricted Web search. Our
aim is to translate automatically as much terms as possible. With this project
we try to demonstrate that having the suitable resources can remarkably
reduce the translation costs.
1 Introduction
Multilingualism is one of the most visible aspects of the change society is living at
present. A new challenge is growing, namely to cope with this change in any sphere
of our lives. In this framework, the major aim of the Linguamón-UOC Chair in
Multilingualism of the Universitat Oberta de Catalunya is promoting a concept of
linguistic diversity that is sustainable, equitable and functional. For that reason, our
team has developed several projects aiming to highlight the importance of linguistic
richness and thus give a chance to all languages, even minor ones.
In this regard, the Linguamón-UOC Chair in Multilingualism has developed a set
of open source tools aiming to automatically extract terminology. Our team has
applied these tools to develop Catalan terminology resources related to the European
Union. Currently, Catalan has an ambiguous status within the EU and very few
documents are translated to this language. One of the reasons that are adduced not to
translate them is the assumption of high costs and the consideration of Catalan as a
minority language. With this project we try to demonstrate that having the suitable
resources and using assisted and automatic translation tools, translation costs are
280
Antoni Oliver and Cristina Borrell
remarkably reduced. Moreover, with a modest budget, the right of Catalan citizens
to have the documentation in their own language can be respected.
As we all know, the European Union produces official documents in 23 official
languages1. The primary law documents must be written in all EU official
languages, in order to respect the transparency, democracy and legitimacy that the
Community promotes. Despite this, many Europeans encounter limitations when it
comes to participating in the European institutions, as much of the multilingual and
multicultural richness that exists within the borders of the EU is not taken into
account.
In order to manage the volume of multilingual documents and assist both citizens
and community workers, the European Union has set up a big amount of specialized
language services2. For instance, the translation service in the European Commission
is the most representative EU linguistic service, and one of the largest translation
services in the world.
The European Commission's Directorate-General for Translation (DGT) has
about 2,500 workers (thus apart from freelance translators) and produces about 1.3
million pages per year. The translation units are organized into departments, one for
each EU language. In recent years, the translation process has been refined,
especially with the incorporation of the technology offered by translation memories,
so that translators can optimize their effort. Roughly, each translator uses to develop
his or her work:
•
•
•
•
Terminology (dictionaries, glossaries, terminological databases, etc.).
Parallel and reference documents;
Access to previously translated texts (especially through translation
memories);
Administration staff responsible for fulfilling the tasks of pre-edition and
post-edition.
Another interesting aspect of the DGT is known as "field offices for
multilingualism", which in some way, are responsible for promoting the richness of
the linguistic diversity of Europe. The function of these branches is to adapt the
Commission's communications to the languages of the region and act as a bridge
between the local citizens and European institutions.
Thanks to the bilateral administrative arrangements between the EU institutions
and the Spanish government3, since June 13th 2005, Catalan citizens may apply to
the institutions in their own language. They may do it in the written communications
to the Council of the European Union and the European Commission, through an
intermediate body designed by the state government in charge of translating the
1
2
3
Bulgarian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek,
Hungarian, Irish, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian,
Slovak, Slovenian, Spanish and Swedish.
All information concerning the languages of Europe and the different language services,
both internal and ‘open’, is available at the website of the EU “Languages and Europe”.
Council conclusions on the official use of other languages (2005 / C 148/01).
IT: Moving Towards Real Multilingualism
281
messages (from Catalan to Spanish), and send the translation to the institution
concerned.
2 Tools
Eurovoc is a multilingual thesaurus covering the areas of activity of the European
Union. Currently, it is available in Bulgarian, Croatian, Czech, Danish, Dutch,
English, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian,
Lithuanian, Polish, Portuguese, Romanian, Slovak, Slovenian, Spanish and Swedish.
It is the only resource that can be downloaded from its website.
Following this information, our team is able to develop a multilingual glossary
with the tools described above. First of all, the team has selected the "available"
languages for the project. English is the most popular language in the EU, followed
by relatively far behind by French and German. The most logical thing then is to
choose English (due to the amount of original documents available) and French,
because a romance language will always be closer to Catalan, our target language,
rather than German.
We will also take Spanish into consideration. This is one of the closest languages
to Catalan, and above all, it is quite easy to find parallel documents in this linguistic
combination (Spanish-Catalan). For instance, and even within the institutional field,
the Official Journal of the Government of Catalonia4 is published both in Catalan
and Spanish.
All that linguistic potential can not be exploited, combined and adapted if we do
not have any tool that helps us handle. That is why we have programmed and used
an automatic extractor of terminology5 that allows us to obtain the equivalent in
Catalan from the Spanish Eurovoc terms. Basically, we have given the Spanish
Eurovoc terms to the official parallel corpus in its Spanish version, thus receiving in
return the official, Catalan version of the term.
Thanks to this collection of tools we can automatically select, manipulate and
retrieve terminology, which is of great help for the translator. It is distributed under
a free software license (GNU/GPL), so the user can freely download it, use it,
distribute it or modify it following his needs. The tools are programmed in Perl, in a
completely modular way, in order to provide, if necessary, its modification and
adaptation. The package works for different platforms, for both Linux and Windows.
This tool allows users to:
•
extract term candidates directly from the text;
•
get help in finding the best translation for a word;
•
automatically build glossaries, both monolingual and multilingual;
•
create terminological databases for specific fields;
•
automatically build lists of lexical units of a specific area.
If the input of the system is a monolingual corpus, the resulting output is a
monolingual list of term candidates. This list must be reviewed manually by a
4
5
DOGC http://www.gencat.net/dogc/
Downloads and more information http://ww.linguoc.cat
282
Antoni Oliver and Cristina Borrell
linguist, who must confirm or reject the proposals automatically selected by the
extractor. Otherwise, if the input consists of a set of documents in multiple
languages, apart from the list we obtain a multilingual terminological resource,
because the system automatically detects any possible equivalent translation of a
candidate.
This tool kit is essentially a linguistic terminology extractor and a device for
automatic translation equivalents search, programmed essentially with statistical
methods.
1. The automatic terminological extractor takes a document, a set of
documents or a parallel corpus and stop−words lists (lists of functional
words) of the languages involved. The statistical method identifies term
candidates based on word frequency in a corpus of expertise. The process,
broadly speaking, is this: the system calculates statistically n-grams
(sequence of n elements, usually from n = 2 to an n determined by the user,
i.e. all combinations of two words, three words etc.). Logically, some of
these combinations will not be relevant, that is why it is essential to filter
the results with the list of empty words. The output of this module will
already be the list of term candidates.
2. In the second case, the system picks a term candidate in a language A and
selects a subset of the parallel corpus with all segments containing this
term. If, then, the system makes a statistical terminological extraction in
these segments, the result should be the term candidate in the language B. It
is expected that the terminological unit more common in the segment in the
language B would be the term that we searched in language A.
3 Procedure
By using the Eurovoc entries in Spanish (6.797 items) and the official, bilingual
Spanish – Catalan parallel corpus we have extracted the equivalents in this
language. The input corpus selected, the Official Journal of the Government of
Catalonia (Diari Oficial de la Generalitat de Catalunya, from now on DOGC), is
published daily from Monday to Friday (except holidays) in two equivalent editions,
one in Catalan and another in Spanish. Basically, the laws, rules and regulations,
general dispositions, agreements, resolutions, edicts, notices, ads and other acts of
Government and its Administration are published in it, both in Catalan and Spanish.
Thanks to it, we have a parallel Catalan-Spanish corpus of institutional, political
and legal areas, that it is why it is one of the Catalan official texts which shares
more characteristics with the EU institutional documents. At the same time, as it
offers a Spanish version, it optimizes the work of the extractor: it simply searches
the Spanish Eurovoc terms in the Spanish version of the DOGC and returns the
Catalan translation for it.
Although the statistical method resolves successfully the search in many cases, it
is not always right. So, the program offers more than one candidate, leaving the user
the final choice. The first result is always the one that has appeared more frequently.
IT: Moving Towards Real Multilingualism
283
Here is where the work really starts for the linguist. Now he or she has to
manually review all the proposals made by the extractor, correct them when they are
incorrect and even suggest translations if the extractor has been unable to offer any.
This will happen in the cases where the Spanish word we are searching for does not
appear in the DOGC. Either way, thanks to the extractor, the linguist saves time and
research.
If the extractor did not act satisfactorily, the expert reviewing the outcome will
have to deal with the types of problems mentioned here6:
1. Often, the first proposal for the Catalan translation is not complete. This
happens especially in cases of terminological units consisting of more than
one word, which somehow mislead statistical systems.
2. Sometimes, Catalan and Spanish versions are not equivalent. This happens
not often, but it can be that the extractor mistakes and confuses the
selection of terms that do not belong to the same field of expertise.
3. Sometimes the extractor proposes n-grams belonging to the same segment
as the key term, but that do not correspond to the possible translation.
4. It can be that versions in other languages other than Spanish do not
correspond to that of Catalan. It is recalled that the extractor takes the word
in Spanish to make the choice of Catalan, so it never takes into account
other languages (in this case, English and French)7.
5. It can also happen that the version in Catalan is the development of an
abbreviation or vice versa. The expert will have to take that into account
and harmonize criteria.
In a second stage, we will translate the terms use different techniques. Firstly,
starting from the terms that are not yet translated, we will generate hypothesis of the
Catalan translation of the Spanish Eurovoc terms, thus by the compositionality
method. Secondly, we will apply statistically automatic translation, and finally we
will check the results in a monolingual corpus or by means of a restricted Web
search. Our aim is to translate automatically as much terms as possible.
4 Results
To perform the experiments we used a parallel corpus Spanish-Catalan (the DOGC)
with all issues from 3.544 (dated 2/01/2002) to 5.118 (dated 24/04/2008). The
corpus has been segmented by sentences and automatically aligned. It consists of a
total of 9 492 333 segments and 121 145 821 words in both languages.
6
7
These errors were found after the implementation of the first version of the extractor,
reason why most of which disappear due to the improvements in later versions. This
analysis helped us discover what changes were important to improve the performance of
the extractor.
Ultimately, this would depend on the quality of the Eurovoc equivalents, which is outside
the scope of this study.
284
Antoni Oliver and Cristina Borrell
The results presented were first obtained for an increment value of n ± 1 and the
search algorithm of translation equivalents was not optimized. Thus the first
candidate to appear was the one that showed the greatest frequency of occurrence in
the parallel subcorpus of all the segments containing the original term. The table
shows the percentage of equivalents appearing in first position, in the first three and
in the top ten positions.
Position
Percentage
1
43,4
1–3
73,4
1 – 10
86,2
Table 1. Results obtained with the algorithm without optimization.
Analysing the results we reached the following conclusions:
•
When candidates get into the top positions with the same frequency, the
candidate who comes in first place can be any of them. This is because
hashes are used to store the candidates, and these data structures do not
have a certain order.
•
Often the first candidate has a higher frequency than the original word. This
means that mono−word terms will appear in the first place even if the
original term was multi−word.
To minimize the consequences of the first phenomenon we programmed an
optimization of the translations' equivalent search algorithm which consisted in
gathering all the proposals that had the highest frequency and apply the algorithm on
the edit distance (the closer the candidate is to the original, the more likely to be the
correct one it is), the results improved as follows:
Position
Percentage
1
62,2
1-3
82,1
1 - 10
88,2
Table 2. Results obtained grouping the more frequent candidates and
using the edit distance.
To avoid the problem produced by candidates with a fake higher frequency we
have applied an optimization similar to the one above, but that including all three
candidates with the highest frequencies. Thus we obtain the following results:
IT: Moving Towards Real Multilingualism
Position
Percentage
1
74,7
1-3
80,4
1 - 10
88,2
285
Table 3. Results obtained by pooling the three candidates with the higher
frequencies and using the edit distance.
As we can see, the results have improved markedly, especially in the top
positions.
Results of the second stage will appear during the second half of the year.
5 Conclusions and future work
In the development of the project we have obtained a multilingual glossary of 2 364
terms used in EU official documents, available in four languages, among which
Catalan. For the moment, above all the official language resources it has only been
possible to process the Eurovoc, which is the only tool that offers the possibility to
be directly downloaded.
During the development we could improve the algorithm three times, in order to
obtain better results. The final result is that the extractor has a 75% of success in the
very first position. The success percentage raises to 88.2% when we look at the first
ten positions: that means that within the first ten choices the translator can find the
correct translation candidate. From the application of the extractor, we obtained a
multilingual glossary of 2 364 terms involving EU policies.
In a second stage, we have used different techniques. Firstly, starting from the
terms that are not yet translated (6 797 – 2 364 = 4 433 terms left), we have
generated hypothesis of the Catalan translation of the Spanish Eurovoc terms, thus
by the compositionality method. Secondly, we have applied statistically automatic
translation, and finally we have checked the results in a monolingual corpus or by
means of a restricted Web search. Our aim is to translate automatically as many
terms as possible.
In sum, we have shown that some basic and simple tools can help us create
multilingual resources that can expedite the process of translating documents. With
the right tools, the process of generation of these resources is very flexible and,
therefore, economic. In addition, all tools used are freely distributed and therefore
available for the whole community8.
This methodology can be applied to a large number of languages. The only
requirements are namely to have a parallel corpus of the desired areas of expertise
and a short list of empty words of the pairs of languages involved. One should keep
in mind that the statistical methodology for the extraction of terminology does not
work well for agglutinative languages.
8
Download and more information about the tools:
http://multilingualismchair.uoc.edu.
286
Antoni Oliver and Cristina Borrell
Future work is divided into two lines:
•
•
On the one hand, to apply the methodology described to other European
terminology resources, such as IATE.
On the other hand, to continue working on improving and optimizing the
extraction tools.
References
[1] European Comission. (2008). Un reto provechoso. Cómo la multiplicidad de
lenguas podría contribuír a la consolidación de Europa [on line]. Brussels.
Internet:
http://ec.europa.eu/education/languages/archive/doc/maalouf/report_es.pdf
[2] Oliver, A., Vazquez, M. & More, J. (2007). Linguoc LexTerm: una
herramienta de extracción automática de terminología gratuita (Tomo 11) [on
line]. Poughkeepsie: Translation Journal, number 4. ISSN 1536-7207.
Internet: http://accurapid.com/journal/42linguoc.htm
[3] European Comission. (2004). Translation and drafting aids in the European
Union languages [on line]. Brussels. Internet:
http://ec.europa.eu/translation/index_en.htm
[4] Diari Oficial de la Generalitat de Catalunya (DOGC) [on line]. Barcelona:
Generalitat de Catalunya, 2008. Internet: http://www.gencat.net/dogc/
[5] Eurovoc. Multilingual Thesaurus [on line]. Brussels: European Union, 2005.
Internet: http://europa.eu/eurovoc/
Introduction of Non-Verbal Means of Communication
in the Corpus of Live Speech*
Tatyana Petrova and Olga Lys
Far Eastern National University, Vladivostok, Russia
Abstract. 1. An integral component of the national corpus is live speech. It
causes some complexity when it is introduced into the corpus. In particular, a
meta-textual annotation which reflects non-verbal components of communication – con-situation and pair-linguistic communicative means such as gestures,
mimics, etc. It is very important to take into consideration the significance of
the non-verbal components for understanding oral texts.
2. Remarks reflecting con-situation are necessary in two types of discourse – in
speech which accompanies actions and in speech which comments events.
3. Remarks for pair-linguistic means, which substitute verbal means of communication are necessary as they are the carriers of information.
Anthropocentrism of new linguistics has caused “change in theoretical priorities”first of all, interest to the discourse, reflecting the whole process of real communication. According to V. A. Plungyan, in the context of a new ideology in linguistics
corpus “has become a powerful instrument of analyzing facts of the language”, which
“has returned to the linguists their real object – texts in a natural language in a maximum volume” [Плунгян 2008]. An integral part of any national corpus is live oral
speech, which makes information about a given language complete. It has been recognized “exclusively actual” for Russian language to create a collection of oral texts as
the corpus researches start on the material of Russian.
Traditionally by live speech is meant an oral speech used spontaneously in
informal communication. It is a special sort of speech – improvisation, created in the
process of the communicative act. It causes some complexity when introducing the
texts of live speech into the corpus. In particular, a meta-textual annotation which
reflects non-verbal components of communication – pair-linguistic communicative
means such as gestures, mimics, looks, activity of the partners. Creators of the first
text-book of conversational Russian [Русская разговорная речь 1978] brought to
notice that sort of difficulty and said that to get an adequate notion of a real conversation, it is necessary to study oral speech in its interaction with non-verbal components
which form communicative act. They also mentioned the necessity of expressing oral
communication in a written form with maximum accuracy. Non-verbal components
play an important part in it: in colloquial linguistics they are con-situation and pairlinguistic means (gestures, mimics, etc.)
*
Researches are made with the support of Russian Humanitarian Scientific Fund (Project
“Creation of the electronic database “Live speech of far-easterners” (№ 08-04-12103в)
288
Tatyana Petrova and Olga Lys
In the National corpus of Russian the phenomena of this kind are expressed by
remarks which in opinion of E. A. Grishina “do not hide any intrigue in themselves”:
they are placed between two “unique marks”(e.g. #All are silent#) and under standard
processing they are automatically transferred to another level of intra-textual existence and this fact lets not consider them when statistic analyses is made [Гришина
2005]. A problem should be admitted with the degree of fulfilling the text with such
remarks. The same problem arises when a multimedia corpus is created (scripts of
films and cartoons, TV- and radio commercials). E.A. Grishina writes: ”If there are no
strict limits, it will bring to the fact when the text of the creators of the corpus will
exceed the text of the film” [Гришина 2005а: 242]. To solve the problem of introducing non-verbal components of communication in the sub-corpus of live speech, in
our opinion, it is necessary we take into consideration the data of colloquial linguistics - how significant are such components for understanding an oral discourse.
Remarks serving to introduce pair-linguistic means and remarks reflecting consituation should be separated. Con-situation is partly demonstrated in meta-description: e.g. The text is characterized by “the place and the time of the events
described”. However the remarks, reflecting con-situational circumstances, which
arise in the process of communication and define the content of the text, are very
important. Speech turns out very con-situational if the topic of it is caused by consituation. Such type of a discourse is a mold of verbal and non-verbal and speech that
accompanies action and speech that comments events refer to it first of all. [Земская
2004]. Examples:
Situation 1: On a commuter train an elderly lady (Grandmother@) and a kid (Alina@)
Alina@ #Moves her hair-brush along the seat# This is a boat floating. #She drops her ball; a
young man picks it up#
Grandmother@ What should you say?
Alina@ Thank you. #She tries to bandage her knee and asks her grandmother# Bandage!
#Grandmother bandages Alina’s knee but she tries to unbind granny’s scarf# Granny,
unbind! #Grandmother is silent.# Unbind! #Grandmother is silent but starts to unbind#
Where is my blood coming from? That’s it; I am going to bind up. Where are my socks?
Grandmother@ #Takes her socks# Put them on!
Alina@ #Looks into the bag# What is in there?
Remarks in the given fragment are necessary since speech, accompanying the
actions of communicators, is only an additional component of their activity and the
purpose of interaction is non-verbal actions.
Situation 2: On a commuter train, women are returning home from the country.
Woman1@ # She sees a jumper# A jumper is coming.
Woman2@ I am about to get off.
Woman3@ #Is looking out of the window# Oh! Second River, it’s getting cooler already.
Woman1@ Second River – what a smell! How people are living here!
Woman2@ #Lookes out of the window# Some pipes.
Woman1@ What pipes! Effluent from all over the city!
Woman3 @ #Lookes out of the window# Fishermen are catching fish.
Woman1@ A then sell that fish, contaminated with helminthes.
Introduction of Non-Verbal Means of Communication in the Corpus of Live Speech
289
The content of the text is determined only by the con-situation – the view from the
window of the commuter train. In this case the use of the remark (#Look out of the
window#) might be possible: the circumstance determines the discourse.
Before saying about introducing pair-linguistic means of communication in the
corpus, it is necessary to explain what it is. The use of mimics, gestures and other
pair-linguistic means is a peculiar feature of a chat. Such means are used “not only to
accompany, to illustrate or enhance the speech but as information carrier as well substitution of a verbal component” [Земская 2004]. As G. E. Kreydlin says, “in the
act of conversation different systems of processing sign information co-exist and they
present a complex proportion” [Крейдлин 2004]. It should be said about a various use
of the term “gesture”: in a narrow notion it is just a gesture, i.e. movement with
hands, with a head, with legs; in a wide notion it is also a movement with a body, postures, expression of the face. Apart from gesticulation all these gestures are the
elements of a communicative act.
All said above lets say about necessity of introducing remarks reflecting pair-linguistic elements of an oral discourse. In our opinion, when organizing the corpus of
live speech, two types of gesture remarks must be differentiated - obligatory and
optional. For this purpose the function of pair-linguistic means should be taken into
consideration with respect to verbal communication. As it is known, there are two
groups of gestures in colloquial linguistics: “sign” – with both a plan of expression
and a plan of content and “non-sign” – with a plan of expression only: they express
rhythm or emotional shadow of an expression [Русская разговорная речь 1983].
Remarks, reflecting rhythmic or emotional gestures, may be considered optional not
obligatory as they don’t have sign function – so important for understanding of a text.
Thus, remarks of the type #Bows his head#; Well, I don’t know! #He makes a helpless gesture”; How could I forget! #She covers her eyes with hands# and so on may
become “unnecessary” for the corpus. As a rule, such gestures double a verbal sign so
it is possible to get along without them.
“Gesture-signs” have a plan of content and transmit information, they can be
demonstrative, descriptive and symbolic [Русская разговорная речь 1983]. The function of substitution is very significant because when it is fulfilled, a non-verbal means
(gesture, mimics, action) become a carrier of information and substitute a verbal component of communication. Gestures of all three types are various and closely interact
with live speech. Nevertheless, it depends how necessary is to use the remarks,
reflecting sign gestures, since a gesture and content of speech interact in various ways
in the boundaries of notional expression.
Gestures-expressions which can be both a motivation and a reaction in a dialogue
should become mandatory elements of short-hand notes introduced into the corpus of
an oral text. Example:
Situation 3:
Woman1@ He is from Artem, isn’t he?
Woman2@ #She shrugs her shoulders#
290
Tatyana Petrova and Olga Lys
The example above includes symbolic gesture “I don’t know” and introduction of
such remark is necessary for keeping the wholeness of the discourse.
Situation 4:
At the seaside, two women – Mother@ and Daughter@
Daughter@ #Goes into the water. Shrinks with cold#
Mother@ What? The water is cold? It’s not time to swim.
Daughter@ Oh, no! It’s just from the beginning, later it will be OK.
In a given fragment a symbolic gesture with a meaning “cold” has a function of a
phrase-motivation which organizes minimal dialogue integrity. However, gestures
from the examples above could accompany verbal expressions as well (e.g. #Goes
into the water. Shrinks# Cold! Or #Shrugs his shoulders# I don’t know). In this case,
gesture frames would be optional, so might be omitted in short-hand notes of the oral
text.
Remarks, denoting demonstrative gestures, and which are tied closely with
demonstrative pronouns might be necessary or mandatory in the discourse in a various degree. Examples:
Situation 5:
Two relatives are in the room – Aunt@ and Niece@.
Auth@ #She thumbs the recipe-book# One more thing. I’ll write down the juice of grapes. And
I’ll make it later.
Niece@ Here is the paper. #She points to the desk# From the wild?
Aunt@ #Bows her head to the cat# Lying. Hypnotizing. Yes, from the wild.
This fragment vividly demonstrates sense integrity of verbal and non-verbal components of communication that is very characteristic of an oral discourse. So the
necessity of the remarks, reflecting both con-situation and demonstrative gestures is
obvious. But gesture remarks of that kind are not obligatory in the cases when a gesture just accentuates a demonstrative pronoun while accompanying it. Example:
Situation 6:
Woman1@ Mushrooms are sold everywhere.
Woman2@ Yes, we’been walking along today. [#Points to the pan with mushrooms#] That’s all
we’ve gathered.
Situation 7:
Junior members of the family (Husband@ and Wife@) tell their grandmother:
Wife@ We went to the beach yesterday. Heaps of crabs there were there!
Husband@ Heaps! They are soft, faded completely. They are so soft!
Wife@ They are so small. [#Demonstrates#]Only one was a palm size. So… So many crabs!
#Makes a gesture “up to the neck”#
In the past phrase it is a relevant remark and it demonstrates a symbolic gesture
meaning “a lot”.
Remarks denoting figurative gestures are not always necessary. They might be
omitted in case the meaning of the gesture is also expressed verbally. Example:
Situation 8:
Two relatives are in the room –Aunt@ and Niece@
Niece@ How is Irishka doing at school?
Introduction of Non-Verbal Means of Communication in the Corpus of Live Speech
291
Aunt@ It’s been just the second week she is at school. She took the same English as yours, like
this. [#Shows with her hands#] Sort of a great book.
Niece@ Ah, by Drozdova.
Thus, in the corpus of live speech obligatory are the remarks to express sign pairlinguistic means which substitute a verbal component of communication and become
a carrier of information. In future an approximate list of the remarks on the basis of
typology developed in the colloquial linguistics might be made.
Solving the problem of introducing non-verbal means of communication in the
corpus of live speech is important not only in respect of “orthographic principle of
representing oral speech”. The system of the remarks, set by definite rules, will be
necessary in case of making the sound-track to the short-hand reports possible. No
doubt, a transcript plays an important part, i.e. is how accurately con-situational circumstances and non-verbal means of communication are expressed, as without them
the whole discourse can’t be understood properly. And it is also obvious that for linguistic research the corpus of live speech may become complete only when it reflects
the picture of real communication in situations of a chat, where we can’t get along
without a non-verbal component.
References
[1] Гришина Е. А. О принципах размещения устных текстов в Национальном
корпусе русского языка // НТИ. Сер. 2. – 2005. – № 3. – С. 31 – 38.
[2] Гришина Е. А. Два новых проекта для национального корпуса:
мультимедийный подкорпус и подкорпус названий // Национальный
корпус русского языка: 2003 – 2005. Результаты и перспективы. – М.:
Индрик, 2005а. – С. 233 – 250.
[3] Земская Е. А. Разновидности городской устной речи и задачи их
изучения // Земская Е. А. Язык как деятельность: Морфема. Слово. Речь. –
М.: Языки славянской культуры, 2004. – С. 247 – 289.
[4] Крейдлин Г. Е. Невербальная семиотика: Язык тела и естественный язык.
– М.: Новое литературное обозрение, 2004.
[5] Плунгян В. А. Корпус как инструмент и как идеология: о некоторых
уроках современной корпусной лингвистики // Русский язык в научном
освещении. – М.: Языки славянской культуры, 2008. – С. 7 – 20.
[6] Русская разговорная речь. Тексты / Отв. ред. Е. А. Земская и Л. А.
Капанадзе. – М.: Наука, 1978.
[7] Русская разговорная речь: Фонетика. Морфология. Лексика. Жест / Отв.
ред. Е. А. Земская. – М.: Наука, 1983.
MorphCon – A Software for Conversion
of Czech Morphological Tagsets
Petr Pořízka1 and Markus Schäfer2
Department of Czech Studies, Faculty of Arts,
Palacký University in Olomouc, Czech Republic
[email protected]
2
Institute of Computer Science
The University of Bonn, Germany
[email protected]
1
Abstract. This study reflects current situation in Czech corpus linguistics with
a special view to morphological annotation of language corpora. Several
morphological tagsets of Czech exist nowadays. These tagsets differ by the
conception reflecting morphological categories in different extent of
complexity. There has also been no possibility of conversion among tagsets.
New tool called MorphCon (Morphological Convertor) is now being developed
for these purposes. This first version (0.1alpha) enables converting of two basic
morphological tagsets of Czech: Prague positional system and Brno’s
attributive system. There are three basic Input/Output (I/O) formats of data
(SimpleTag-Conversion, KWIC/Tag-Format, WPL-Format) within version
0.1alpha. Tagsets are implemented into the MorphCon as "drivers" with
"encode" and "decode" function as well as an "universal library" called DZInterset (© Daniel Zeman) – modified in our tool – plays key role for the
process of conversion as a transcoder. The MorphCon software is thus built as
an universal converter: modularity, the Interset as a transcoder, possibility of
adding of another tagsets (not only Czech ones) and I/O formats.
1 Introduction
The development of the MorphCon software (shortened from Morphological
Convertor), an application based on formal rules and algorithms, was motivated by
the current situation of Czech corpus linguistics with respect to morphological
annotation of linguistic corpora. At present several morphological tagsets exist for
Czech, which reflect morphological categories to various degrees of complexity and
which differ one from another in their conceptions. The predominant and most used
system is the one designed by J. Hajič (a positional tagset, hereafter, PT) [6][7], as
seen in the written part of the Czech National Corpus [18]. Not less important is the
Brno system of morphological tags (an attributive tagset, hereafter, AT), the
conception of which is authored by K. Osolsobě [11] and R. Sedláček [17]. This
system is made use of by the morphological analyzer (tagger) called AJKA [15][16]
in the corpora of the Natural Language Processing Centre at Masaryk University,
Brno (hereafter, NLP FI MU).
MorphCon – A Software for Conversion of Czech Morphological Tagsets
293
Among others belongs Petkevič´s morphological tagset used in the international
project MULTEXT-EAST (the Orwell 1984 Corpus) [13] or, most recently, the socalled kódovník (coder), used lately for tagging of the Prague Spoken Corpus [4].
The tagsets mutually differ with respect to their conception, the degree of
complexity and system as far as the coded grammatical categories are concerned.
Morphological tagsets also depend on particular software applications, which are not
composed to annotate linguistic data by means of another tagset, as they operate with
one particular tagset only. However, if necessary, it is possible to annotate the same
text by various tagsets by means of the respective applications, but the crucial role is
played here by the efficacy, i.e. the degree of success and failure at (semi)automatic
processing of the texts by the given taggers, the scope and structuring of their
dictionary for lemmatization and tagging etc.
One of the possible solutions to this situation is a software allowing automatic
conversion of (Czech) morphological tagging systems. For these reasons, the
MorphCon converter has been developed, which allows the conversion of corpus data
already tagged by one tagset into a different tagset.
The conversion between the Prague and Brno tagsets has been chosen for the first
phase of this software. The Prague system is a positional one (15 positions are given,
to each of which a particular linguistic category is assigned, represented by a defined
subset of symbols), the Brno system is an attributive one (there is a combination of
diagrams, in which the first symbol represents a grammatical category, the second one
its concrete value for the word given). A detailed account and an overview of the
symbols of both systems are presented in Hajič [6][7], Osolsobě [11], and Sedláček
[15][16][17]. For us to find out the possibilities of the mutual automatic conversion, it
was necessary to carry out an analysis and comparison of both tagsets, the key factors
being mutual compatibility, loss of information and other aspects (see below).
2 The MorphCon converter
The software MorphCon (v0.1alpha) is in development since 2008 by a team of
authors consisting of members of three universities: Petr Pořízka (Faculty of Arts,
Palacký University, Olomouc), Marek Schäfer (Faculty of Informatics, Bonn
University, Germany) and Daniel Zeman (Institute of Formal and Applied Linguistics,
Faculty of Mathematics and Physics, Charles University, Prague). It is developed in
the Perl scripting language (v5.10.0) and designed as a universal converter, the
fundamental elements of which are modularity and existence of a conversional tagset
(which excludes the direct conversion). This allows implementation of further tagsets
and Input/Output formats.
With respect to possible future users-linguists, the so-called Graphic User
Interface (GUI) has been kept in mind.
294
Petr Pořízka and Markus Schäfer
Fig.1. The Graphic User Interface of the MorphCon converter
2.1 The structure of the software
The conception, structure and the principles of the MorphCon converter are
presented in the Scheme in the Supplement, to which the following text refers to. The
MorphCon is composed of several components/modules (all written in the Perl
scripting language):
• GUI: the Graphic User Interface – MorphCon.pl
• Input/Output modules: MorphCon::{simple,kwic,wpl}
• Drivers: implemented morphological tagsets –
tagset::cs::{attributive-ajka,pdt,positional-16}
• Universal library: a modified DZ Interset library
The key quality is modularity of the software, apart from offering universality
(including the possibility to enrich the MorphCon with other tagsets) also variability
in the process of conversion, i.e. various/different settings of the input and output
data. The MorphCon is based on the universal tagset DZ Interset (see [21] [22]) which
works as a transducer when converting one tagset into another. Each tagset is
implemented into the software as a „driver“ with double function, either as a source,
or target tagset:
• encode-function: source-tagset → Interset
• decode-function: Interset → target-tagset
2.1.1 The Interset
The Interset functions as a „feature-projection“, i.e. it is structured as
a „feature::value“ system, the grammatical, morphosyntactic category being the
feature. It must also contain all the features (grammatical categories) with their values
from all implemented tagsets. An overview of all features and drivers may be found
on the websites of the project [24]. During the conversion, the tags from the tagset A
MorphCon – A Software for Conversion of Czech Morphological Tagsets
295
are converted into the Interset, from which they are subsequently converted into the
target tags within the tagset B. The quality of conversion thus depends on the quality
of algorithms converting the particular categories with values of the tagsets given into
the “feature::value” system of the DZ Interset [23].
Example: The Interset and the “feature::value” structure for the attributive tag
'k2eAgInPc6d1' from the ajka-tagset in the Perl script:
{
'animateness' => 'inan',
'case' => 'loc',
'degree' => 'pos',
'gender' => 'masc',
'negativeness' => 'pos',
'number' => 'plu',
'pos' => 'adj',
'possgender' => 'masc',
'possnumber' => 'plu',
'tagset' => 'cs::ajka'
}
Annotation of the tag 'k2eAgInPc6d1'
k2
eA
gI
nP
c6
d1
k = part of speech
e = negation
g = gender
n = number
c = case
d = degree
2 = adjective
A = affirmation
I = masculine inanimate
P = plural
6 = sixth
1 = positive
The principle of conversion of this tag from an attributive one into a positional
tagset with representation in the Perl script in the "AT,PT {Interset} format" looks as
follows:
Example:
k2eAgInPc6d1,AAIP6MP--1A-----{'animateness' => 'inan','case'
=> 'loc','degree' => 'pos','gender' => 'masc','negativeness'
=> 'pos','number' => 'plu','pos' => 'adj','possgender' =>
'masc','possnumber' => 'plu','tagset' => 'cs::attributiveajka'}
2.1.2 The input/output modules
The Input/Output modules of the MorphCon noticeably extend the options of the DZ
Interset, since – apart from the process of conversion itself – they allow to set the
input and output data variably. Thus, the input and output formats do not have to be
identical.
296
Petr Pořízka and Markus Schäfer
The options for the Input/Output modules
–
–
–
the data file
• the .txt format, "plain text"
the tag format
• the Brno tagset – attributive:
• the Prague tagset – positional:
tagset::cs::attributive-ajka
tagset::cs::pdt
tagset::cs::positional-16
the file format
• simple: SimpleTag-Conversion
• KWIC: KWIC/Tag-Format
• WPL: WPL-Format
Note on the data file
So far, the MorphCon has only been working with the plain text (.txt files), but in
the future it should also process texts with structured data format (SGML and
especially XML data).
Note on the tag format
There are two variants for the positional tagset: (a) "tagset::cs::pdt" is the original
driver of the Interset, in which the category of aspect is not included; (b)
"tagset::cs::positional-16" comes from the original driver, but the important category
of aspect has been newly added to it in the MorphCon. (This means an extension from
15 to 16 tagset positions.) Therefore the positional system of tags may be enriched by
means of conversion in the direction AT → PT, as the attributive system counts with
the aspect category, while the positional system does not (apart from some
exceptions). It is a way to prevent a loss of information during conversion in such a
case when the source tagset contains a grammatical category not found in the target
tagset.
Note on the file format
As mentioned above, the MorphCon allows the change of input and output file
formats. When converting tagset A into tagset B, it is thus possible to determine the
resulting data format, differing from the source data format A. So far the MorphCon
has been working with three formats: (1) The SimpleTag-Conversion (shortened:
simple), (2) The KWIC/Tag-Format (shortened: KWIC), and (3) The WPL-Format
(shortened: WPL). See the following examples of these formats:
2.1.2.1 The SimpleTag-Conversion
This format is a simple conversion from tagset A into tagset B. One separate line
corresponds to one tag of the respective tagset:
MorphCon – A Software for Conversion of Czech Morphological Tagsets
AT
k3gInPc6
k2eAgInPc6d1
k1gInPc6
297
PT
PDXP6----------AAFP6----1A----NNIP6-----A-----
Table 1. The simple "tag-to-tag" conversion
2.1.2.2 The KWIC/Tag-Format
The KWIC/Tag-Format allows the conversion of tags from already annotated and
completed corpora. In practice, the so-called concordances are most frequently
searched for, i.e. search results where the key word (KWIC = Key Word In Context)
is surrounded by a context (see the example taken from the Bonito corpus manager –
for more detailed information about the Bonito concordancer see Rychlý [14]).
Usually the tag is placed as a metatext following the word, this means exactly in the
format KWIC/Tag (the slash is a division mark). In the converting process, only tags
are converted, the remaining text is kept unchanged.
The source tagset (AT) – the pre-conversion phase
172 měl zajištěnou nominaci na Světový pohár v < Petrohradu /k1gInSc6wS > . " Má
Švanda naději startovat
284 prezident Václav Havel za přítomnosti dalších < hostí /k1gMnPc2wK > . Za zvuku
státní hymny vztyčili
↓
The target tagset (PT) – the post-conversion phase
172 měl zajištěnou nominaci na Světový pohár v < Petrohradu /NNIS2-----A----- > . "
Má Švanda naději startovat
284 prezident Václav Havel za přítomnosti dalších < hostí /NNMP2-----A---2- > . Za
zvuku státní hymny vztyčili
2.1.2.3 The WPL-Format
The corpus data are usually structured into the so-called vertical format, there is
always only one word (WPL = Word Per Line) with its linguistic interpretation per
line. For that reason, the sequence word – lemma – tag is most often separated by
a tabulator (comma or other marks are acceptable, too). The WPL-Format format is
therefore very important in case we build a corpus: we annotate the data with the help
of tagset A, but we may subsequently need to convert the data into tagset B. The
reason for this may be situations when a tagger with implemented tagset A has at its
disposal a higher-quality and larger dictionary and gives higher-quality results (i.e.
proportionally fewer errors during the process of automatic annotation) than a tagger
with implemented tagset B. In comparison with tagset A, tagset B is more userfriendly, may be remembered more easily etc. There may be a whole variety of
reasons for using this format.
298
Petr Pořízka and Markus Schäfer
AT: word
lemma
tag
PT: word
lemma
tag
V
těch
dlouhých
rozhovorech
v
ten
dlouhý
rozhovor
k7
k3gInPc6
k2eAgInPc6d1
k1gInPc6
V
těch
dlouhých
rozhovorech
v
ten
dlouhý
rozhovor
RR--6----------PDXP6----------AAFP6----1A----NNIP6-----A-----
Table 2. The sequence word – lemma – tag in the WPL-Format
3 Different conceptions of AT and PT systems
Important facts for conversion, i.e. conversion algorithms, are the question of
different conceptions of tagsets and their mutual convertibility, the potential loss of
information during conversion, and other aspects. There may be cases where
modifications of tagsets are concerned, see the example of a positional tagset supplied
with another position representing the category of aspect.
Problems when converting the attributive and positional tagsets of Czech
appeared, in fact, we still find ourselves at the stage of testing and adjusting the
conversion algorithms to optimize the MorphCon conversion process.
The positional tagset is characterized by the controversial second position
(SUBPOS), which is a rather unsystematic and heterogeneous mixture of symbols
from various grammatical categories of various parts of speech and with various
degrees of complexity (cf. [7]). The second position contains 75 values altogether,
which according to our opinion could be reduced, avoiding loss of relevant linguistic
information. The „double-category“ of the 3rd/6th and 4th/7th positions (gender and
number), and existence of the 6th and 7th categories as separate positions expressing
possessive gender and possessive number, respectively, are also debatable. On the
other hand, the grammatically important category of aspect is missing. The category
of aspect has been indeed added as the 16th position to the annotated corpora of Czech
called SYN2005 and SYN2006PUB (cf. [2][3]), but it has not been implemented into
the positional tagset (and tagger). The question remains, what the still unexploited
„reserved“ positions 13 and 14 are intended for, and why the tagset still contains these
two empty positions. The attributive tagset is more complex concerning the contained
linguistic (sub)categories and could be characterized as the one that is more
systematic regarding the conception, more user-friendly regarding acquisition and
more easily remembered. The problem that appeared when working on the conversion
algorithms concerned documentation and a complete set of tags. There exist several
versions of the attributive tagset from various development stages, but there is no
version history available, and therefore we had to take all of them into account. There
exist four versions of the attributive tagset in total: (1) the official version by R.
Sedláček (the author of Ajka, the tagger using the attributive tagset), available on the
website of NLP FI MU [17]; (2) the tagset table by R. Sedláček from 2006; (3) the
tagset version found in the disambiguation manual by Bartůšková et al. [1]; (4) the
original tagset version from R. Sedláček´s diploma thesis [15].
MorphCon – A Software for Conversion of Czech Morphological Tagsets
299
A number of Perl scripts were created for testing of the MorphCon, which had
partial functions during the conversion process, e.g. indicating tags of the same
content and visualization of differences between "encode" and "decode" functions of
"drivers", i.e. tagsets. With help of testing scripts, the problematic situations and
errors originating before the conversion, namely already during the process of
annotation of corpus data, may be detected. Mainly non-uniformed tag order,
erroneous annotations of particular tokens, which remained in the text most probably
due to insufficient disambiguation or were inserted into the tags by a disambiguator
(the „human factor“), e.g. an attribute without a value or a value without an attribute,
more values in the category of case and other errors.
4 Further advancement of the MorphCon
In the near future, we plan to expand functionality of the MorphCon. We intend to
implement other tagsets of Czech (the above mentioned mte-cz tagset or kódovník),
further Input/Output formats (both linear and vertical), e.g. the format used by the
Ajka tagger [15] or in the Prague Dependency Treebank (the so-called csts /pml/
format – for more details, see PDT 2.0 Guide, Chapter 3. Data) [12].
Following the experience acquired during the development of the MorphCon and
solving problems relating to the conversion, implementation of some scripts originally
intended for testing purposes only were considered. The aim is to produce a new
module „Tag-Checker“, which would serve as a helping instrument for
disambiguation and tag checking (checking their correctness/error rate, anomalies
etc.). Basic information about the MorphCon software may be found at
http://www.morphcon.webnode.cz.
References
[1]
[2]
[3]
[4]
[5]
Bartůšková, D., Hlaváčková, D., Ungermanová, M. (2004). Manuál pro
značkování a desambiguaci slovních tvarů v jazykových korpusech. Brno.
[online], [cit. 2009-05-29]. Available from:
<http://nlp.fi.muni.cz/projekty/desman/desman1603.pdf>
Corpus SYN2005. [online], [cit. 2009-05-29]. Available from:
<http://www.korpus.cz/english/syn2005.php>
Corpus SYN2006PUB. [online], [cit. 2009-05-29]. Available from:
<http://www.korpus.cz/english/syn2006pub.php>
Čermák, F. et al. (2007). Frekvenční slovník mluvené češtiny. Praha:
Karolinum.
Hajič, J., Hladká, B. (1998). Tagging Inflective Languages: Prediction of
Morphological Categories for a Rich, Structured Tagset, In: Proceedings of the
Conference COLING – ACL `98. Montreal, Canada, pp. 483–490.
300
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
Petr Pořízka and Markus Schäfer
Hajič, J. (2005a). Popis morfologických značek – poziční systém. In: Manuál
korpusového manažeru Bonito. [online], [cit. 2009-05-29]. Available from:
<http://www.korpus.cz/bonito/znacky.php>
Hajič, J. et al. (2005b). A Manual for Morphological Annotation – Positional
Tags. [online], [cit. 2009-05-29]. Available from:
<http://ufal.mff.cuni.cz/pdt2.0/doc/manuals/en/m-layer/html/ch02s02s01.html>
Hladká, B. (2000). Czech Language Tagging. PhD Thesis. Charles University,
Prague.
MorphCon – Convertor of Czech Morphological Tagsets. [online], [cit. 200905-29]. Available from: >http://morphcon.webnode.cz>
Natural Language Processing Centre (Faculty of Informatics, Masaryk
University, Brno). [online], [cit. 2009-05-29]. Available from:
<http://nlp.fi.muni.cz/en/nlplab>
Osolsobě, K. (1996). Algoritmický popis české formální morfologie a strojový
slovník češtiny (unpublished PhD Thesis). Brno.
PDT 2.0 Guide. [online], [cit. 2009-05-29]. Available from:
<http://ufal.mff.cuni.cz/pdt2.0/doc/pdt-guide/en/html/index.html>
Petkevič, V. Popis morfologických značek použitých v korpusu orw-mte.
[online], [cit. 2009-05-29]. Available from:
<http://ucnk.ff.cuni.cz/orwell_znacky.php>
Rychlý, P. (2000). Korpusové manažery a jejich efektivní implementace. (PhD
Thesis) Brno. [online], [cit. 2009-05-29]. Available from:
<http://www.fi.muni.cz/~pary/dis.pdf>
Sedláček, R. (1999). Morfologický analyzátor češtiny. (MA Thesis) Brno.
[online], [cit. 2009-05-29]. Available from:
<http://nlp.fi.muni.cz/projekty/ajka/ajka.pdf>
Sedláček, R. (2004). Morfologický analyzátor češtiny Ajka. [online], [cit.
2009-05-29]. Available from: <http://nlp.fi.muni.cz/projekty/ajka/ajkacz.htm>
Sedláček, R. (2006). AJKA tagset. Brno. [online], [cit. 2009-05-29]. Available
from: <http://nlp.fi.muni.cz/projekty/ajka/tags.pdf >
The Czech National Corpus. [online], [cit. 2009-05-29]. Available from:
<http://www.korpus.cz/english/index.php>
The Prague Dependency Treebank. [online], [cit. 2009-05-29]. Available from:
<http://ufal.mff.cuni.cz/pdt2.0/>
ÚFAL Tools (Institute of Formal and Applied Linguistics at the Faculty of
Mathematics and Physics, Charles University, Prague). [online], [cit. 2009-0529]. Available from: <http://ufal.mff.cuni.cz/tools.html>
MorphCon – A Software for Conversion of Czech Morphological Tagsets
[21]
[22]
[23]
[24]
301
Zeman, D. (2008). Reusable Tagset Conversion Using Tagset Drivers. In
Proceedings of the Language Resources and Evaluation Conference, LREC
2008. CD full edition + printed Conference Abstracts. Marrakech, Morocco.
[online], [cit. 2009-05-29]. Available from:
<http://ufal.mff.cuni.cz/~zeman/publikace/2008-02/tagdrivers-marrakech-styllrec.pdf>
Zeman, D. DZ Interset. [online], [cit. 2009-05-29]. Available from:
<https://wiki.ufal.ms.mff.cuni.cz/user:zeman:interset>
Zeman, D. DZ Interset – Features. [online], [cit. 2009-05-29]. Available from:
<https://wiki.ufal.ms.mff.cuni.cz/user:zeman:interset:features>
Zeman, D. DZ Interset – Tag Set Drivers. [online], [cit. 2009-05-29].
Available from: <https://wiki.ufal.ms.mff.cuni.cz/user:zeman:interset:drivers>
Recent Developments in
the National Corpus of Polish
Adam Przepiórkowski1,5 , Rafał L. Górski2 , Marek Łaziński3,5 , and Piotr P˛ezik4
1
2
Institute of Computer Science, Polish Academy of Sciences, Poland
Institute of Polish Language, Polish Academy of Sciences, Poland
3 Polish Scientific Publishers PWN, Poland
4 University of Łódź, Poland
5 University of Warsaw, Poland
Abstract. The aim of the paper is to present recent – as of July 2009 – developments in the construction of the National Corpus of Polish. The main developments are: 1) the design of text encoding XML schemata for various levels of
linguistic information, 2) a new tool for manual annotation at various levels, 3)
numerous improvements in search tools.
1
Introduction
The aim of the paper is to present recent – as of July 2009 – developments in the
construction of the National Corpus of Polish (Pol. Narodowy Korpus J˛ezyka Polskiego;
NKJP; http://nkjp.pl/). The project, funded by the Polish Ministry of Science and
Higher Education (project number: R17 003 03), was launched at the very end of 2007
and will run until the end of 2010.
The rest of this section presents the background of the project, the consortium, and
envisaged applications of the results of the project in the humanities. The following
section, §2, describes the expected results of the project in general terms.6 Section 3
presents recent developments concerning corpus annotation and search tools, as well as
text encoding standards deployed in the project. Section 4 concludes the paper.
1.1
Background
For Polish, the most represented Slavic language of the EU, there still does not exist a
national corpus, i.e., a large, balanced and publicly available corpus, which would be at
least morphosyntactically annotated. Currently, there exist three contemporary7 Polish
corpora which are – to various extents – publicly available. The largest and the only one
that is fully morphosyntactically annotated is the IPI PAN Corpus (http://korpus.
pl/; Przepiórkowski 2004), containing over 250 million segments (over 200 million
6
7
These two sections overlap to a large extent with Przepiórkowski et al. 2008.
Another – much smaller and dated, but historically very important – corpus is available in
its entirety: http://www.mimuw.edu.pl/polszczyzna/pl196x/, a 0.5-million word corpus
created in the 1960. as the empirical basis of a frequency dictionary (Kurcz et al. 1990).
National Corpus of Polish
303
orthographic words), but – as a whole – it is rather badly balanced.8 Another corpus,
which is considered to be carefully balanced, the PWN Corpus of Polish (http://
korpus.pwn.pl/), contains over 100 million words, of which only 7,5 million sample
is freely available for search. The third corpus, the PELCRA Corpus of Polish (http:
//korpus.ia.uni.lodz.pl/), also contains about 100 million words, all of which are
publicly searchable.
1.2
Consortium
The project is unique in that it involves all major corpus developers for a given language,
including the three developers of the three corpora mentioned above: the Institute of
Computer Science at the Polish Academy of Sciences (ICS PAS; the Polish acronym
is IPI PAN, hence the name of the corpus) in Warsaw, which coordinates the project,
the PWN Publisher in Warsaw and the PELCRA group at the University of Łódź. The
fourth partner is the Institute of Polish Language at the Polish Academy of Sciences
(IPL PAS) in Cracow, the developer of an internal corpus, available only for research
carried out at the Institute.
Another institution whose staff and students are involved in the project is the University of Warsaw. Some of the ideas which influenced the methodology of the project
evolved in the Faculty of Polish Studies and in the Institute of Informatics, where two
of the authors lecture on corpus linguistics, linguistic engineering and Polish grammar.
1.3
Practical applications in the humanities
The project is correlated with another national project, led by IPL PAS, aiming at the
development of a new large dictionary of Polish, for which the National Corpus of
Polish will serve as the empirical basis.
External users, such as lexicographers and linguists, will mainly be interested in
searching the corpus for word and phrase concordances as well as collocations. The
corpus can also serve as the treasure of well-known quotations from Polish and key
words of Polish culture, with some emphasis on good representation of secondary
school required readings in Polish literature and history. Therefore, quotations from the
corpus will be crucial for new large dictionaries of Polish (including the new dictionary
currently developed at IPL PAS, as well as dictionaries published by PWN), not only
as a source of the typical uses of words, but also as a reference to cultural authorities
rooted very well in the Polish literary tradition.
The corpus as a whole also enables creating a number of comparable corpora. The
size of the corpus, exceeding the informal standard of 100 million, shall guarantee that
there will be sufficient number of texts of different genres to meet the selection criteria
of target corpora. Technically, there are two possible ways of creating a comparable
corpus: to create a separate subcorpus comparable to a corpus of a different language or
to place in the header of selected texts an element stating that this very text is a part of a
comparable corpus. Hence, the user will have the option of narrowing down their query
8
There exists a 30-million segment subcorpus of the IPI PAN Corpus which is relatively balanced.
304
Adam Przepiórkowski et al.
to the texts marked as included in the comparable corpus, in the same way as they can
narrow down the query to a certain author, genre or period.
A subproject of the NKJP consists in monitoring the words in daily and weekly
newspapers and comparing word frequencies in two periods of time in order to test
the saliency of words in every day public discourse. A demo version of this tool has
recently been made available on the web site of the Association of Local Press (http://
www.gazetylokalne.pl/). The project “Words of the Week” will be launched soon in
cooperation with the weekly “Polityka” and later with Polish dailies. (A similar project
was in progress on the website of the daily Rzeczpospolita from 2004 to 2007, see
Łaziński and Szewczyk 2006).
2
Corpus
The intended size of the whole National Corpus of Polish is 1 billion words, of which
at least 300-million word subcorpus will be carefully balanced.
2.1
Representativeness
In establishing the criteria of representativeness we build on our own experience, as well
as on the experience of other national corpora, especially, the Czech National Corpus.
We understand representativeness as representing the perception of language by a certain linguistic community (cf. Čermák et al. 1997), what in practice means reflecting
the structure of readership in the structure of the entire corpus. There are theoretical
reasons for this choice, besides practical considerations. Out of potential concepts of
representativeness which may be applied to various corpora, the following two have the
best methodological motivation: representing the population of texts or representing
the structure of readership. If we adopted the model representing the production of
text, around 90% of the corpus would consist of press. Hence the corpus would be
representative but not balanced. The use of such a corpus in linguistic and lexicographic
work is questionable (Górski 2008). The representativeness is based on several pools
exploring the choices of reading of the average Polish reader as well as the circulation
of the press. On the other hand we expect the corpus to be not only representative but
also balanced, therefore the amount of press is lowered compared to the overall picture
of the readership, so as not to let any text type dominate over the entire corpus (Górski
2009).
A first step towards establishing representativeness is determining the typology of
text types. The typology is generally based on the bulk of work on Polish stylistics.
There are however some text types which seem to be overseen by traditional stylistics,
which had to be added. The classification is based mainly on intralingusitic features of
texts. A small pilot study was conducted, so as to establish a set of purely linguistic
factors differentiating the text types (Górski and Łaziński 2008). We are however aware
of the fact, that the proposed classification is not the only possible.
As assigning a text to a certain type is not always a straightforward task, we decided
that every text will be classified in a double-blind procedure by two linguists. In case of
discrepancy they will have to discuss their decision.
National Corpus of Polish
305
To assure a wide coverage by topic we use classifications used by the libraries
which will be encoded in the header of each text. To meet the needs of philologists
and lexicographers we shall try as far as possible try to include in the corpus the most
important works of modern Polish literature including poetry.
2.2
Spoken component
A 30 million word component of the NKJP will represent the spoken register of Polish.
This part of the project is coordinated by the PELCRA team and it derives from the
experience of compiling the 600 000 spoken-conversational component of the PELCRA
corpus (Waliński and P˛ezik 2007). Apart from transcripts of public speeches, parliamentary commission proceedings, televised debates, chat shows, radio interviews and
news bulletins, a 3 million word subset of the spoken component will comprise natural,
spontaneous conversations recorded by persons trained to preserve the natural character
of the language data collected. Spoken NKJP data will be annotated with sociolinguistic
metadata, including information on the age, gender, education and social background
of the recorded speakers. Selected fragments of the spoken corpus will be aligned with
the recordings and integrated in a relational database engine on top of which a publicly
accessible web interface will be implemented (P˛ezik et al. 2004).
2.3
Annotation
The entire corpus will be annotated linguistically, structurally and with bibliographic
metadata. The basis of the linguistic annotation will be the full morphosyntactic annotation (not just parts of speech, but also values of cases, genders, etc., as appropriate). As
in the IPI PAN Corpus, each segment (token) will contain not only the information about
which morphosyntactic interpretation is correct in a given context, but also about all
the other possible interpretations, rejected in the context (Przepiórkowski et al. 2004).
Apart from the morphosyntactic annotation, the corpus will contain partial syntactic
information, i.e., main types of syntactic groups will be identified, as well as named
entities and various kinds of lexical constructions (so-called syntactic words). Also,
forms of over 100 frequent semantically ambiguous lexemes will be disambiguated.
Because of the size of the corpus, it will not be possible to annotate the whole corpus manually. However, a 1-million word subcorpus of the representative 300-million
subcorpus, reflecting its structure, is being annotated manually and it will be utilised for
training and testing of linguistic tools which will subsequently be used for the automatic
annotation of the whole corpus.
3
3.1
Recent developments
Text encoding
The need for text encoding standards for language resources (LRs) is widely acknowledged: within the International Standards Organization (ISO) Technical Committee 37
/ Subcommittee 4 (TC 37 / SC 4; http://www.tc37sc4.org/) work in this area has
306
Adam Przepiórkowski et al.
been going on since early 2000s, and working groups devoted to this issue have been
set up in two current pan-European projects, CLARIN (http://www.clarin.eu/) and
FLaReNet (http://www.flarenet.eu/). It is obvious that standards are necessary for
the interoperability of tools and for the facilitation of data exchange between projects,
but they are needed also within projects, especially where multiple partners and multiple
levels of linguistic data are involved.
NKJP is committed to following current standards and best practices, but it turns
out that the choice of text encoding for multiple layers of linguistic annotation is far
from clear. Przepiórkowski and Bański 2009 contains an overview of recent and current
standards and best practices, at various stages of development. The conclusion of this
paper is that the guidelines of the Text Encoding Initiative (Burnard and Bauman 2008;
http://www.tei-c.org/) should be followed, as it is a mature and carefully maintained de facto standard with a rich user base. Various proposed standards proposed by
ISO TC 37 / SC 4, including Morphosyntactic Annotation Framework (MAF), Syntactic
Annotation Framework (SynAF) and Linguistic Annotation Framework (LAF), are still
under development and, especially SynAF and LAF, have the form of general models
rather than specific off-the-shelf solutions.
Nevertheless, when selecting from the rich toolbox provided by TEI, an attempt has
been made to follow recommendations of these proposed ISO standards, as well as other
common XML formats, including TIGER-XML (Mengel and Lezius 2000) and PAULA
(Dipper 2005). This work, described in more detail in Przepiórkowski and Bański 2009
(cf. also Bański and Przepiórkowski 2009), resulted in TEI P5 XML schemata encoding
data models largely isomorphic with or at least mappable to those formats.
3.2
Annotation tools
As mentioned in §2.3 above, a 1-million word subcorpus of NKJP is being annotated
manually and it will be used for training automatic annotation tools. Anotatornia, a tool
for the manual annotation of word senses developed within a previous project at ICS
PAS (Hajnicz et al. 2008), has been extended and extensively modified to allow for
the manual addition of sentence boundaries, word-level segmentation, morphosyntactic
annotation and word sense disambiguation (Przepiórkowski and Murzynowski 2009).
At each level, annotation is adduced by two linguists, connecting with the server via
a web interface. In case of differences, both are notified that the other annotator made
a different decision at a given place, but they are not informed about each other’s decision. Each annotator may change their own annotation, or confirm it. If the discrepancy
persists, a referee makes the final decision and suggests a modification of the annotation
guidelines (Przepiórkowski 2009b), if necessary.
Currently annotation is performed at the levels of sentence segmentation, word-level
segmentation and morphosyntax, with word sense disambiguation expected to start in
August 2009.
For the manual morphosyntactic annotation, each segment is automatically marked
with all interpretations known to the new version of Morfeusz (Woliński 2006), a morphosyntactic dictionary of Polish based on the data of the Słownik gramatyczny j˛ezyka
polskiego (‘Grammatical dictionary of Polish’; Saloni et al. 2007). The task of the
annotator is to select the right interpretation or add the correct interpretation, if it is not
National Corpus of Polish
307
among those proposed by Morfeusz. In the process, various deficiencies of Morfeusz
and the underlying grammatical data base have been identified and corrected.
The morphosyntactic tagset used in NKJP is a modified version of the IPI PAN
Tagset (Przepiórkowski and Woliński 2003a,b). The differences between the two tagsets
are described – and a formal specification of the NKJP Tagset presented – in Przepiórkowski 2009a.
3.3
Search tools and NKJP demo
Since developing an efficient search tool able to manage a 1-billion corpus is a potentially high-risk task, two approaches are pursued in parallel.
The first approach is based on the combination of Apache Lucene (http://lucene.
apache.org/) and relational database technologies, and it is partly inspired by the
implementation of the PELCRA Corpus of Polish: we expect this approach to scale well
with the size of the corpus. Apart from scalability, this search engine also focuses on
providing convenient access to concordance and collocation search results in a variety
of output formats, including downloadable spreadsheets, compressed URL-s, integrated
browser plugins and web services. It has yet to be seen to what extent this approach will
accommodate more complex types of linguistic search at various levels of annotation.
The second approach is based on Poliqarp (Janus and Przepiórkowski 2007a,b), a
dedicated search engine developed at ICS PAS and currently serving a corpus of 250
million segments: while Poliqarp involves a very expressive query language, currently
further expanded to accommodate syntactic queries, it is not clear how well it scales
with the size of the corpus. So far, modifications of Poliqarp within NKJP consisted in
developing a new corpus compiler, translating the TEI-based XML encoding of texts to
an efficient binary format. End-user improvements include more specific error messages
in case of not well formed queries, and an option to randomise search results.
Both search engines are successfully employed in the NKJP Demo (http://nkjp.
pl/index.php?page=6&lang=1), which currently consists of about 500 million words.
4
Conclusion
As any “recent developments” publication, this paper describes work in progress. Intensive work within the National Corpus of Polish project concerns all levels of corpus
development: from data acquisition, through text encoding and linguistic annotation, to
efficient corpus search engines. We hope this overview paper has whetted the reader’s
appetite for the final project results.
References
Bański, P. and Przepiórkowski, A. (2009). Stand-off TEI annotation: the case of the National Corpus of Polish. In Proceedings of the Third Linguistic Annotation Workshop
(LAW III) at ACL-IJCNLP 2009, Singapore.
Burnard, L. and Bauman, S., editors (2008). TEI P5: Guidelines for Electronic Text
Encoding and Interchange. Oxford. http://www.tei-c.org/Guidelines/P5/.
308
Adam Przepiórkowski et al.
Čermák, F., Králík, J., and Kučera, K. (1997). Recepce současné češtiny a reprezentativnost korpusu. Slovo a slovesnost, 58, 117–124.
Dipper, S. (2005). Stand-off representation and exploitation of multi-level linguistic
annotation. In Proceedings of Berliner XML Tage 2005 (BXML 2005), pages 39–50,
Berlin.
Górski, R. L. (2008). Representativeness of the written component of a large reference
corpus of Polish. Primary notes. Forthcoming.
Górski, R. L. (2009). The representativeness of NKJP. Talk delivered at Practical
Applications in Language and Computers (PALC 2009), Łódź, April 2009.
Górski, R. L. and Łaziński, M. (2008). Wzór stylu i wzór na styl. Zróżnicowanie
stylistyczne tekstów Narodowego Korpusu J˛ezyka Polskiego. Talk delivered at the
VII Forum Kultury Słowa, Gdańsk, October 2008.
Hajnicz, E., Murzynowski, G., and Woliński, M. (2008). ANOTATORNIA – lingwistyczna baza danych. In Materiały V konferencji naukowej InfoBazy 2008, Systemy
* Aplikacje * Usługi, pages 168–173, Gdańsk. Centrum Informatyczne TASK, Politechnika Gdańska.
Janus, D. and Przepiórkowski, A. (2007a). Poliqarp 1.0: Some technical aspects of a
linguistic search engine for large corpora. In J. Waliński, K. Kredens, and S. GoźdźRoszkowski, editors, The proceedings of Practical Applications in Language and
Computers PALC 2005, Frankfurt am Main. Peter Lang.
Janus, D. and Przepiórkowski, A. (2007b). Poliqarp: An open source corpus indexer
and search engine with syntactic extensions. In Proceedings of the ACL 2007 Demo
and Poster Sessions, pages 85–88, Prague.
Kurcz, I., Lewicki, A., Sambor, J., Szafran, K., and Woronczak, J. (1990). Słownik
frekwencyjny polszczyzny współczesnej. Wydawnictwo Instytutu J˛ezyka Polskiego
PAN, Cracow.
Łaziński, M. and Szewczyk, M. (2006). Słowa klucze w semantyce i statystyce. słowa
tygodnia „Rzeczpospolitej”. Biuletyn Polskiego Towarzystwa J˛ezykoznawczego,
LXII, 57–68.
Mengel, A. and Lezius, W. (2000). An XML-based encoding format for syntactically
annotated corpora. In Proceedings of the Second International Conference on Language Resources and Evaluation, LREC 2000, pages 121–126, Athens. ELRA.
P˛ezik, P., Levin, E., and Uzar, R. (2004). Developing relational databases for corpus
linguistics. In B. Lewandowska-Tomaszczyk, editor, The proceedings of Practical
Applications in Language and Computers PALC 2003, Frankfurt am Main. Peter
Lang.
Przepiórkowski, A. (2004). The IPI PAN Corpus: Preliminary version. Institute of
Computer Science, Polish Academy of Sciences, Warsaw.
Przepiórkowski, A. (2009a). A comparison of two morphosyntactic tagsets of Polish.
In Proceedings of the Mondilex workshop in Warsaw, June 2009.
Przepiórkowski, A. (2009b). Zasady znakowania morfosyntaktycznego w NKJP. Version 1.19 of 27 July 2009.
Przepiórkowski, A. and Bański, P. (2009). Which XML standards for multilevel corpus
annotation? Unpublished manuscript.
Przepiórkowski, A. and Murzynowski, G. (2009). Manual annotation of the National
Corpus of Polish with Anotatornia. Talk delivered at Practical Applications in Language and Computers (PALC 2009), Łódź, April 2009.
National Corpus of Polish
309
Przepiórkowski, A. and Woliński, M. (2003a). A flexemic tagset for Polish. In Proceedings of Morphological Processing of Slavic Languages, EACL 2003, pages 33–40,
Budapest.
Przepiórkowski, A. and Woliński, M. (2003b). The unbearable lightness of tagging:
A case study in morphosyntactic tagging of Polish. In Proceedings of the 4th International Workshop on Linguistically Interpreted Corpora (LINC-03), EACL 2003,
pages 109–116.
Przepiórkowski, A., Krynicki, Z., D˛ebowski, Ł., Woliński, M., Janus, D., and Bański,
P. (2004). A search tool for corpora with positional tagsets and ambiguities. In
Proceedings of the Fourth International Conference on Language Resources and
Evaluation, LREC 2004, pages 1235–1238, Lisbon. ELRA.
Przepiórkowski, A., Górski, R. L., Lewandowska-Tomaszczyk, B., and Łaziński, M.
(2008). Towards the National Corpus of Polish. In Proceedings of the Sixth International Conference on Language Resources and Evaluation, LREC 2008, Marrakech.
ELRA.
Saloni, Z., Gruszczyński, W., Woliński, M., and Wołosz, R. (2007). Słownik gramatyczny j˛ezyka polskiego. Wiedza Powszechna, Warsaw.
Waliński, J. and P˛ezik, P. (2007). Web access interface to the PELCRA referential
corpus of Polish. In J. Waliński, K. Kredens, and S. Goźdź-Roszkowski, editors,
The proceedings of Practical Applications in Language and Computers PALC 2005,
pages 65–86, Frankfurt am Main. Peter Lang.
Woliński, M. (2006). Morfeusz — a practical tool for the morphological analysis of
Polish. In M. A. Kłopotek, S. T. Wierzchoń, and K. Trojanowski, editors, Intelligent
Information Processing and Web Mining, Advances in Soft Computing, pages 511–
520. Springer-Verlag, Berlin.
Spoken Texts Representation in the Russian National
Corpus: Spoken and Accentologic Sub-Corpora
Svetlana Savchuk
Vinogradov Institute of Russian Language of the Russian Academy of Sciences
Moscow, Russia
[email protected]
Abstract. The paper presents the main characteristics of the Spoken and
Accentologic corpora, available on open access on the site ruscorpora.ru.
Corpora size, type and sources of text material, the way it is represented in the
corpora, types of linguistic annotation, corpora composition and ways of their
effective use according to their purposes are being described.
Keywords: spoken corpora, spoken data representation, linguistic annotation,
accentology, the Russian National Corpus
1 Introduction
Modern corpora as a rule contain a spoken component, because it is oral speech that
forms the core of linguistic performance. The interest in investigation of spoken data
is constantly increasing in corpus linguistics, as referred to [1]. There are common
recommendations and standards concerning the preparation and representation of oral
texts in the corpus (TEI, EAGLES). Meanwhile each national corpus (British [2],
Czech [3], Slovak [4], Russian [5], American [6] and so on) offers its own way of
texts selection and corpus architecture.
As far as the Russian National Corpus is concerned, it has two distinctive features
in organization of the spoken component. 1) Spoken corpus is excluded from the main
body of written texts and forms a separate module in the RNC supplied with the types
of annotation specific for spoken texts only. 2) We have refused the idea of creating a
“universal” spoken corpus suitable for all kinds of oral speech study – from phonetic
aspects to discourse analyses. That’s why there is not one but three spoken subcorpora within the RNC, with their own features and spheres of application. These are
a) Spoken sub-corpus, b) Accentologic sub-corpus (both are functioning) and c)
Multimedia Corpus (is under development).
The paper presents the main parameters of the Spoken and Accentologic corpora.
2 The Corpus of Spoken Russian
The Corpus of Spoken Russian is being created in cooperation with different research
centers: Moscow, Saint Petersburg, Saratov, Perm, Yekaterinburg, Omsk, Krasnoyarsk, Uljanovsk, etc. Nowadays it accumulates both transcriptions of present-day
spoken speech and data collected and described by the researchers since the 1960s.
Spoken Texts Representation in the Russian National Corpus
311
The Corpus of spoken texts which functions as a part of the RNC provides researchers
with ample opportunities.
2.1 Criteria for text selection
1) The Spoken corpus includes full authentic texts. It enables us to discover
everything which could have been missed if we studied only selective records.
2) The corpus includes an amount of texts which significantly exceeds the average
number of texts at the disposal of a colloquial speech researcher. It makes possible to
explore frequency and randomness of a phenomenon, discover regular patterns which
could only be found within a huge amount of texts and draw statistically reliable
conclusions from these patterns.
3) The corpus contains texts heterogeneous in terms of when and where the records
were made. They also differ in regard to sex- and age-related, social and professional
features of speakers. It should be mentioned that dialectal texts are not included in the
Spoken corpus, they are collected in the Dialectal corpus and supplied with inherent
annotation.
4) The texts in the spoken corpus belong to a long time span of more than 70 years
(if we take the scripts of the 1930s movies into account). The first records of spoken
language date from 1956, the latest were made in spring 2009. It enables us to follow
changes in colloquial speech (which are very swift), note any new trends, etc.
5) The corpus (in contrast to collections normally used for colloquial speech
studies) includes spoken texts from different spheres of communication, which have
been pronounced under different circumstances. We do not share some researchers’
opinion that only ‘natural speech of city residents in direct contact’ may be considered
as ‘real Russian speech’[7]. Spoken speech as a form of a language’s existence (as
opposed to written form) is presented in different spheres of functioning: in everyday
life sphere – as private (non-public) spoken speech; in educational and scientific
sphere – as academic spoken speech; in journalism – as public spoken speech, TV
speech and radio speech; in official sphere – as official spoken speech; in industrial
and technical sphere – as professional spoken speech; in theological sphere – as a
sermon; in advertising sphere – as TV and radio advertisements; in fiction – as speech
of theatre and cinema. That is why a spoken text in the corpus is not only a dialogue
in a shop or a family conversation at dinner table, but also an academic lecture, a
report at a seminar, an author’s meeting with readers, an interview or a TV talk-show,
sports commentaries and some.
6) Another criterion for distinguishing different types of spoken speech, which is used
for selecting corpus texts, is a level of preparation or spontaneity. We can use this scale to
place different kinds of spoken texts in decreasing order in terms of spontaneity. The
extreme points on this scale are spontaneous, non-prepared speech (dialogue and
monologue) on the one hand and written-to-be-spoken speech (reading a text aloud,
pronouncing a text learned by heart) on the other hand. There are a lot of text types
between these points which were called quasi-spontaneous [8]: interview, narrative on
312
Svetlana Savchuk
foreknown topic, speech prepared according to premeditated plan, retelling of another
person’s speech, formulaic speech, etc. In the corpus there are no spoken texts which
would represent written-to-be-spoken speech. But there are a lot of texts which would be
estimated as quasi-spontaneous on this scale. First of all, these are the records of public
speech and the movie sub-corpus.
7) The movie sub-corpus includes scripts of live-action movies and animated films
(and in future also documentaries and commercials). It is a unique component of the
Spoken sub-corpus in the RNC. This language sphere until recently has been escaping
spoken language researchers’ and corpora authors’ attention. Meanwhile, the
influence of these texts on Russian (and not only Russian) language usage is great, as
it has been shown in [9].
2.2 Composition and structure of the Spoken corpus
At present the spoken corpus consists of 8.6 mln tokens. It may be considered as a
representative text collection, which reflects the functioning of contemporary Russian
spoken language. Let us show what effect it has on the corpus composition and
structure1. The distribution of spoken texts according to main parameters is shown in
tables below.
Sphere
Public speech
Private speech
Speech of cinema
Tokens
4392123
780150
3378096
Percentage
51
9
40
Table 1. The distribution of texts according to spheres of spoken communication
Sphere
Public
speech
Private
speech
Text types
discussion
talk
interview
lecture
sports commentary
narration
parliamentary hearings
conference
round table
other
conversation
telephone conversation
tale
retelling
microdialogues
argument
Tokens
1920306
1232823
309918
195302
182811
112820
86640
48972
49177
253354
609167
80648
47340
12533
25435
5027
Percentage
43.7
28.1
7.1
4.4
4.2
2.6
2
1.1
1.1
5.7
78.1
10.3
6.1
1.6
3.3
0.6
Table 2. The distribution of spoken texts according to text types
1
Data are given as of July 2009.
Spoken Texts Representation in the Russian National Corpus
Sphere
Speech of
cinema
Text types
drama
comedy
detective
films for children
costume films
science fiction and fantasy
action
animated films
other
Tokens
844713
1194060
313611
258368
112487
93156
57823
102882
400996
313
Percentage
25.0
35.3
9.3
7.6
3.3
2.8
1.7
3
12.0
Table 3. The distribution of speech in movies according to film genres
The corpus includes texts of different topics. The texts with the mark ‘private life’
are the most frequent (more than 50% of all texts). Politics and social life texts, art
and culture texts, scientific texts, leisure and entertainment texts, sport texts are less
frequent in decreasing order.
As for the time of recording, the majority of texts dates from 2003 to 2008, the
other significant part of the texts (more than 600 thousand tokens) are from the 1990s,
947 thousand from 1980s, about 1,9 mln from 1960-1970s, more than 420 thousand –
before 1960 (the last are mainly movie transcripts) .
Period
1930–1959
1960–1979
1980–1989
1990–1999
2000–2008
Percentage
5
22
10
8
55
Table 4. The distribution of spoken texts according to date of recording
Geographical distribution of the spoken corpus is quite broad. The corpus
includes texts recorded in Moscow and Moscow region (the majority of texts), in
Saint Petersburg, Kirov region, Saratov, Samara, Taganrog, Voronezh, Novosibirsk,
Uljanovsk, Yekaterinburg.
2.3 Language data presentation method
Texts for the Spoken corpus are derived from various sources. A considerable part of
them are records of spoken texts published in reading books and collections compiled
by the researchers of spoken language and edited by E. A. Zemskaja, O. A. Lapteva,
N. N. Rozanova and M. V. Kitaigorodskaja, A. S. Gerd, O. B. Sirotinina and others in
the 1970-1990s. Another part are records of spoken texts collected in such research
centers as Vinogradov Institute of Russian language, Moscow State University, Saint
Petersburg State University, Saratov University, Uljanovsk University. Transcripts of
sociologists’ conversations in focus groups, based on different important social topics
314
Svetlana Savchuk
form an important part of public spoken speech in the Corpus. These records were
rendered by the Public Opinion Foundation. All movie transcripts and a high
percentage of records of spoken texts are made by the members of the RNC team or
under their supervision.
All texts are represented in modern orthography. The syntagm boundaries are
marked by slashes. Punctuation marks (dot, elision, interrogation and exclamation
marks) are used for marking the end of a phrase. Some typical features of oral speech
are reflected in transcripts. This involves an increase in word-forms with non-standard
spelling. The issue is dealt with preliminary orthographical normalization of
transcripts. All cases of substandard spelling are supplied with their standard
equivalents which are recognized and annotated later with grammar features attributed
to the whole complex. As a result on screen display we can see the peculiarities of
pronunciation, captured by non-standard word-forms (поскоку, тыща in the example
below) with the correct grammar parsing (because the normalized forms поскольку,
тысяча are analyzed). You would get the same if you input these words in the corpus
dictionary and had them identified correctly by the lemmatizer.
[Васильев, муж, 1962] но вот смотрим / Аня Бога́лий / она в общем-то
недалеко / она рядышком / как я уже говорил / основна… основные события
будут развиваться на огневом рубеже / поскоку / здесь уже я повторяюсь / но
для тех / кто не слышал / здесь в Сан-Сика́рио очень сложная высота /
практически почти тыща семьсот метров над уровнем моря / и дышится ну
крайне тяжело / очень сильно разря́женный воздух [Константин Выборнов,
Дмитрий Ваcильев. Спортивный репортаж: биатлон. Олимпийские игры.
Эстафета 4х6 км Женщины (23.02.2006) // «Первый канал», прямой эфир, 2006].
2.4 Types of linguistic annotation
The same types of annotations as in the RNC are typical of the Spoken corpus –
metatextual, morphological and semantic. Thereby, the same types of subcorpora and
the same types of search as in ‘the written corpus’ are available in the spoken corpus.
Metatextual annotation marks a text as a whole and includes information regarding
author’s name, sex, age or date of birth, text characteristics (creation date, sphere of
functioning, text type, movie genre, domain) and so on [10].
Morphological information is assigned to a word-form and consists of four groups
of tags: 1) lexeme (a dictionary form of the lexeme and the part of speech to which it
belongs); 2) a variety of the lexeme's grammatical features, known as wordclassifying features; 3) a variety of the word-form's grammatical features, known as
word-altering features; 4) information concerning non-standard forms of the wordform, orthographic variations, etc.
Sociological annotation is specific to the spoken corpora only. It is assigned to
different speaker’s utterances and characterizes a word usage from the point of view
of sex and age of a speaker (if this information is available). Sociological annotation
Spoken Texts Representation in the Russian National Corpus
315
allows a user to create his/her own sub-corpora by various parameters or their
combinations: by a speaker’s sex (so a user could create a sub-corpus of feminine or
masculine spoken language); by a speaker’s age (for example, a user can create a subcorpus of teenagers’ phrases); by a speaker’s year of birth (this option is available
only for movie transcripts, so you could select the phrases by the actors born in 19 th
century); by an actor’s name (for example, you can create a sub-corpus of Eugene
Leonov’s phrases).
Apparently, sociological annotation may be supplemented with metatextual
annotation which makes it possible to select texts by one speaker and include his/her
name and year of birth in the description of the text. It is clear, that if a) there are
more than one speaker, b) speakers cannot be named because of ethical reasons, c)
their age is unknown or speakers are of very different age, this information cannot be
included in the description of the text. In this case all we can refer to is sociological
annotation.
2.5 Prospects of development
As it has been mentioned above, spoken language in the RNC is presented in three
corpora. In terms of data representation the spoken corpus is apparently at a
disadvantage comparing to both the accentologic corpus (because it has information
about stress) and the multimedia corpus (because it portraits real language). But still,
does it have any prospects of development? We should answer ‘yes’ to this question
because of several reasons.
First of all, the spoken corpus differs from the accentologic and multimedia
corpora in text types. As we have already mentioned, there are spoken texts recorded
in different regions of Russia at different times. In fact, if we had records, money and
staff nothing could deter us from matching transcripts with audios and process the
texts the same way as in the accentologic corpus.
However, it is not always possible. The significant part of spoken texts (first of all,
earlier records and collections from regional research centers given to the RNC) exist
only as transcripts. Tape records either have been lost or have never been made (if
microdialogues were recorded manually). In the first place it applies to the texts
published in reading books [11], [12], [13]. These records can only be presented as a
part of the spoken corpus.
In spite of the improvement of equipment making transcripts of audio records is
still the most widespread and the most reliable way to put down spoken texts. This
resource shouldn’t be underestimated. As the experience of practical training in
collecting spoken texts by Moscow students has shown, transcripts are not always
supplemented with satisfying audio records which could be used in the corpus. There
are different reasons for this – technical or just accidental: poor quality of the records,
rare file formats, conversion errors, etc. Such records cannot be used in the
accentologic corpus, but they can take place in the spoken corpus. Thereby, the
316
Svetlana Savchuk
spoken corpus excels both the accentologic and upcoming multimedia corpora in the
number and diversity of texts.
The second reason for continuing development of the spoken corpus is the nature
of linguistic annotation and the corpus search engine. The corpus is an effective
research tool only if its annotation corresponds to the linguistic goals defined by a
researcher. Thus, researches of most morphological, syntactic and lexical features of
spoken language are easier to conduct using the materials of the spoken corpus.
Among its main advantages are its size, diversity of texts, annotation similar to the
one in the written corpus, which allows to compare results easily. When researching
phonetic, accentologic, prosodic and paralinguistic features of spoken utterances, it is
better to refer to the accentologic or multimedia corpora.
Thus, the most urgent problem of the spoken corpus development is increasing of
its size up to 10 million tokens. This can be achieved by balancing the corpus and
adding texts which are not yet very well presented – private speech in the first place.
Another goal is to broaden the geography of the corpus by adding Russian spoken
texts from different regions, the near abroad and foreign countries. This would enable
us to study the Russian language which functions in contact with related, non-related
and foreign languages.
3 The Accentologic corpus of Russian
Accentologic corpus gives an opportunity to study word stress. This information is
very important for languages with non-fixed stress. The Russian language is one of
them. It has a very complicated accentual system. Russian stress has the following
features: firstly, it is non-fixed, which means that any syllable may be stressed (for
example, зо’лото, воро’на, борода’); secondly, it is variable, which means that it
may change its place as a result of inflexion or word formation (for example, зо’лото
– золото’й, золоти’ть, позоло’та; борода’ (n, f, nom, sg) – бороды’ (n, f, gen,
sg), бо’роду (n, f, acc, sg), бо’роды (n, f, nom, pl), боро’дка (diminutive form).
Moreover, the Russian stress system is in process of rearrangement, when
significant changes are rather swift. That is why from the very beginning the
Accentologic corpus was planned as diachronic, which would allow studying the
history of the Russian stress. Meanwhile, there are no limitations of the size of the
corpus (except technical capabilities). The goals of the corpus define its structure,
composition and criteria for text selection.
3.1 Criteria for text selection
There are two zones in the Accentologic corpus.
1) The zone of prose includes oral texts and films transcripts, in which stressed
syllables are marked according to real pronunciation. The main criterion for including
a text in the corpus is the availability of a record of sufficient quality which would
allow us to verify the transcript. We are equally interested in both accentologic
Spoken Texts Representation in the Russian National Corpus
317
standards of the literary language and their variants which change over a period of
time, because text annotation makes it possible to correlate any variant to a sphere of
functioning, a genre and a speaker. Thus, the prose zone contains some examples of
spontaneous everyday speech, spoken public speech of different levels of spontaneity
(TV and radio speech, political speeches, academic spoken speech, sermons, etc.),
movie and radio play transcripts, reading aloud. The earliest records of this zone date
from the beginning and the first decades of the 20th century (gramophone records of
L.N. Tolstoy’s letters, political leaders’ speeches, records of speeches made at the
First congress of writers in 1934, films of 1930s). In perspective some of accented
written texts (for example, books and manuscripts of the 18th and 19th centuries, and
later – even older texts) may be included in this zone.
2) The zone of poetry contains texts with marked accented syllables, so one can
define the exact word stress using special rules. Specially annotated poetic texts of the
18th-20th centuries are included in this zone and still continue to be added. At present
this zone mainly reflects the history of the Russian stress, because the corpus contains
poetry written before the 20th century.
3.2 Structure and composition of the corpus
Nowadays the Accentologic corpus contains more than 6.4 million tokens. Texts’
distribution among the two zones and according to time periods is listed in the tables
below.
Zone
Poetry
Prose
Speech of cinema
Public speech
Private speech
Reading aloud
Tokens
3238956
3198998
not available
not available
not available
Percentage
50.5
49.7
Table 5. Texts proportion in the zones of the Accentologic corpus
Zone
1700 -1799
1800-1910
1911-1949
1950-1979
1980-1999
2000-2008
Poetry
738148
2500808
-
Prose
248237
1537243
865503
548015
Table 6. Distribution of texts according to creation date
3.3 Types of annotation
The Accentologic corpus is supplied with four types of annotation which are used in
the Spoken corpus, and besides that has its own accentologic mark-up. In metatextual
annotation not only characteristics common to both zones of prose and poetry are
318
Svetlana Savchuk
used (the author’s name, age, sex, date of text recording/ creating), but also
parameters intrinsic to each zone severally. These are genre, meter, clausula, rhyme,
line size for poetry and text type for prose.
Each word in Accentologic Corpus has morphological and semantic annotation as
in the RNC in whole. Sociological annotation which attributes the information about a
speaker’s sex, age (or date of birth) to each utterance is used in the zone of prose only.
Due to accentologic annotation each word is supplied with stress marks, so you
can make different kinds of search requests and retrieve data of stressed or unstressed
word-forms in combination with grammatical and semantic features.
3.4 Language data preparation and presentation method
Preparation of texts for Accentologic corpus is performed in several steps.
The first step includes audio files decoding, orthographical normalization and
transcripts editing. The second stage is automatic and uses the software Accentuator
(made by A. Polyakov) which accent words in accordance with the recommendations
of the built-in lexicon. This lexicon is compiled from the database of normative
Russian dictionaries, but is also amplified by the corpus developers. At the third stage
an expert listens to audio records, reviews them against transcripts and corrects
annotated texts. As a result we get a text accented the same way it is pronounced.
[Сергей Ненашев, Александр Абдулов, муж, 38, 1953] Ро́вно в четы́ре де́сять вы́
вклю́чите телеви́зор / настро́ив его́ на четвёртый кана́л. [Виктор Сергеев, Игорь
Агеев. Гений, к/ф (1991)]
As we have mentioned before, in the zone of poetry the same annotation as in the
Poetry corpus is used. The special programme annotates up beats and then an expert
corrects any mistakes according to the special manual. As a result we get a text which
looks like this:
Он идѐт в ворота̀, он ужѐ на крыльцѐ, Он взошѐл по круты̀м ступеня̀м На
площа̀дку и вѝдит: с печа̀лью в лицѐ Одино̀ко-уны̀лая та̀м [В.А. Жуковский.
Иванов вечер (1822)]
This example is trisyllabic anapaest in which the marked up beats (ictuses) match
the real word stress.
Вот на̀ш геро̀й подъѐхал к сѐням; Швейца̀ра мѝмо о̀н стрело̀й Взлетѐл по
мра̀морны̀м ступѐням, Распра̀вил во̀лоса̀ руко̀й, Вошѐл. [А. С. Пушкин. Евгений
Онегин / Глава первая (1823-1824)]
In this iambic tetrameter the number of up beats (ictuses) is bigger than the
number of word stresses. In this case we should exclude all impossible stresses (which
are not presented in any dictionaries or reference books) and take into consideration
possible stresses only. For the word-form мра’морным the only possible stress is the
stress on the first syllable, for the word-form волоса’ (n, m, nom, pl) – the stress on
the ending, but the stress on the first syllable would characterize this word-form as
gen. sg. Interestingly enough, in these two examples the stresses of the form dat. pl. of
the noun ступеням are different, which shows the coexistence of these variants in
the beginning of the 19th century.
Spoken Texts Representation in the Russian National Corpus
319
3.5 Intended use and prospects of development
The Accentologic corpus is one of specialized corpora and aims for researches in a
quite narrow sphere of the Russian accentology. In spite of rather small size of the
corpus it is still enough to give a researcher an opportunity to study new material,
correct some points of his/her theory and normative recommendations [14].
Furthermore, the corpus enables us to solve methodological problems: it can be
used as a reference material, as a material for compiling exercised and teaching aids
for those learning Russian.
In the nearest future the increase of corpus size and text variety is planned.
According to this plan poems of the 1st half of the 20th century will be included in the
zone of poetry. The zone of prose will be expanded due to texts belonging to different
spheres of spoken communication and created in various time periods. Among them
are academic lectures, interviews and TV talk-shows, sports commentaries, sermons,
political speeches, narratives, private conversations and so on. Many of these texts
will be available on-line next year.
References
[1] Rayson, P., Mariani, J. Visualizing corpus linguistics. In: Corpus Linguistics
2009. 20-23 July 2009. Abstracts, p. 201. Liverpool (2009).
[2] British National Corpus,
http://www.natcorp.ox.ac.uk/corpus/creating.xml.ID=design#plan
[3] Czech National Corpus, http://ucnk.ff.cuni.cz/english/struktura.php
[4] Slovak National Corpus, http://korpus.juls.savba.sk/structure/index.en.html
[5] Russian National Corpus, http://ruscorpora.ru/search-spoken.html
[6] American National Corpus, http://www.americannationalcorpus.org/
SecondRelease/ contents.html
[7] Matveeva, T. V. (ed). Zhivaja rech uralskogo goroda. Teksty (Живая речь
уральского города. Тексты). Yekaterinburg Univ. Press, Yekaterinburg
(1995)
[8] Galyashina, E.A. The differentiation features of prepared and spontaneous
discourse in oral and written speech (Проблема дифференциации
спонтанной и подготовленной речи) In: Papers from the Annual
International Conference “Dialogue 2002”, http://www.dialog-21.ru/materials/
archive.asp?id=7287 &y=2002&vol=6077
[9] Grishina, E. A. Dva novych projekta dl’a Natsional’nogo korpusa:
multimedijnyj podkorpus i podkorpus nazvanij (Два новых проекта для
Национального корпуса: мультимедийный подкорпус и подкорпус
названий). In: Natsionalnyj korpus pusskogo jazyka: 2003–2005. Rezul’taty i
perspektivy. Indrik, Moscow (2005)
320
Svetlana Savchuk
[10] Grishina, E. A. Ustnaja rech v Natsional’nom korpuse russkogo jazyka
(Устная речь в Национальном корпусе русского языка) In: Natsional’nyj korpus pusskogo jazyka: 2003–2005. Rezul’taty i perspektivy. Indrik,
Moscow (2005)
[11] Zemskaya, E. A., Kapanadze, L. A. (eds.) Russkaja razgovornaja rech: Teksty
(Русская разговорная речь: Тексты). Nauka, Moscow (1978)
[12] Kitajgorodskaya, M. V., Rozanova, N. N. Rech moskvichej: Kommunikativnokulturologicheskij aspect. (Речь москвичей: Коммуникативнокультурологический аспект). 2nd ed. Nauchnyj mir, Moscow (2005)
[13] Sergieva, N. S., Gerd, A. S. (eds.) Russkaja razgovornaja rech jevropejskogo
severo-vostoka Rossii (Русская разговорная речь европейского северовостока России). Syktyvkar (1998)
[14] Grishina, E. A. Natsional’nyj korpus russkogo jazyka kak istochnik svedenij
ob ustnoj rechi. In: Rechevyje tekhnologii, 3. (2009)
The Meaning of the Conditional Mood Within
the Tectogrammatical Annotation of Prague
Dependency Treebank 2.0
Magda Ševčíková
Institute of Formal and Applied Linguistics
Faculty of Mathematics and Physics
Charles University in Prague, Czech Republic
Abstract. The conditional form is one of the moods of Czech verbs, and it renders several meanings in contemporary Czech texts (Sect. 2). The present paper
focuses on the primary function of this mood, which is to express hypothetical
events (Sect. 3). In Section 4, we briefly mention how modality has been treated
up to now in PDT 2.0 and some other treebanks and finally in Section 5 we
propose a new way how the primary meaning of the conditional mood should be
captured in the annotation scheme of the tectogrammatical layer of PDT 2.0.
1
Introduction
Verbal mood as a universal morphological category occurs in many languages of the
world, however, it is structured differently in different languages (see, e.g., [1], [10]).
In Czech, three moods are traditionally distinguished: indicative, imperative, and conditional. In this paper, attention is focused on the Czech conditional mood, in particular on
its primary function. As the corpus data document, the conditional is primarily used to
refer to events which may be generally characterized as hypothetical; more specifically,
to events which are contingent on realization of other events, to events which cannot
happen any more etc. We want to demonstrate that in contemporary Czech the conditional as a means expressing hypotheticality of an event is a semantically relevant means
which contributes to the meaning of the whole sentence. Thus, it has to be included in
the representation of the meaning of the sentence.
In Prague Dependency Treebank 2.0 (PDT 2.0 in the sequel), the linguistic meaning
of the sentence is represented as a tectogrammatical tree, which consists of nodes and
edges with a set of attributes (see [6]). However, neither in PDT 2.0 nor within Functional Generative Description (FGD in the sequel; c.f. [14]), on the basis of which the
annotation scenario of PDT 2.0 was built, any considerable attention has been paid to
the semantics of the morphological category of mood.
After an overview of the functions of the conditional mood (Section 2), the semantic
relevance of the conditional in its primary function is documented in Section 3 of the
present paper. In Section 4, the treatment of the conditional and other modal means in
PDT 2.0 (and in FGD) as well as in some other treebanks is described. A new formal
means (called grammateme in PDT 2.0 and FGD) for capturing the primary function
of the conditional within the tectogrammatical annotation of PDT 2.0 is proposed in
Section 5. The paper concludes with some final remarks in Section 6.
322
2
Magda Ševčíková
Functions of the conditional mood
2.1
The primary function of the conditional
The morphological category of verbal mood is acknowledged to be one of the means by
which modality is expressed in Czech. Modal meanings are further expressed by many
other means such as modal verbs, modal particles, prosody etc. The primary function
of the conditional is described under several terms in Czech linguistic literature. In
this paper, we use the term hypotheticality to refer to the semantics of this mood. In
its primary function, the conditional is opposed to the indicative by means of which
events are simply asserted, presented as given. In Grammar of Czech (see [8], [2]),
the semantic opposition of hypotheticality vs. assertion is called factual modality. The
conditional mood is considered to be the marked member in this opposition.
Two types of hypothetical events are further distinguished: events which could
happen (potential events) vs. events which cannot happen (irreal events).1 Potential
events are expressed by the so called present conditional as one of two subcategories
of the conditional mood. The second subcategory of the conditional, the so called past
conditional, expresses irreal events unambiguously. However, the present conditional
is frequently (or even predominantly) used instead of the past conditional in Czech in
the last decades. Such a substitution has been regarded as acceptable if the irreality is
signalized by an adverb etc. (e.g. in [5]). Nevertheless, in the analyzed corpus data, the
present conditional was used as an expression of irreal events also in cases in which the
irreality could be resolved only on the basis of a very broad context or of knowledge of
situation. Such an ambiguity causes substantial problems with annotation.
2.2
Secondary functions of the conditional
Besides the primary function, the conditional fulfills also other (secondary) functions in
Czech. For example, in the sentences Doporučil bych vám tu smlouvu podepsat ‘I would
recommend you to sign the contract’ or Uvedl byste konkrétní příklad? ‘Would you give
a concrete example?’, the speaker uses the conditional to express his recommendation
or appeal in a more polite way than using the indicative.
The conditional can further express the speaker’s will to do something (c.f. Kouřil
bych ‘I would smoke’, which actually means “I want to smoke”) or his conviction
whether the content will be realized or not (e.g., Šel bych ‘I would go’, which can
be used instead of “I will probably go” in an appropriate context). This paper concerns
with the primary function of the conditional; contexts in which it is used in its secondary
functions are not analyzed here.
1
The term ‘irreal’ is used according to [17], i.e., it refers to events which are impossible rather
than to unreal events.
The Meaning of the Conditional in PDT 2.0
3
323
The conditional mood as a means for expressing hypothetical
events
3.1
Corpus material analyzed
The primary function of the conditional mood was studied on the basis of language data
from two corpora: from PDT 2.0 (http://ufal.mff.cuni.cz/pdt2.0; see [6]) and
from SYN2005 corpus (http://ucnk.ff.cuni.cz).
PDT 2.0 is a collection of contemporary Czech newspaper texts to which a morphological annotation and annotation at two syntactic layers was assigned, at the so
called analytical layer (layer reflecting the surface shape of the sentence) and at the
tectogrammatical layer (layer of the linguistic meaning of the sentence). Annotation of
all three types is available for more than 49 thousand sentences (i.e., more than 830
thousand tokens); the train data set (about 80 % of the full data) was used to search for
sentences to be examined.
SYN2005 is a representative corpus of contemporary written Czech which contains
100 million tokens. In comparison with PDT 2.0, only morphological annotation was
assigned to the data of SYN2005.
3.2
Semantic relevance of the conditional
In order to show the semantic relevance of the conditional mood in its primary function,
we proposed a substitution test. In the test, the conditional, which is supposed to be
the marked member of the opposition of the factual modality, was substituted for the
indicative as for the unmarked member of this opposition. The substitution test was
performed in several types of sentences from PDT 2.0 and SYN2005 corpus in the
following way: in a concrete sentence which involved a present or past conditional
verb form, the conditional was replaced by the indicative. Then, we tested whether the
resulting sentence with the indicative can be used as an expression of the hypothetical
event in question and whether it is acceptable with regard to the immediate as well as
broader context, to our knowledge of situation etc. (c.f. [16] for more details on the
substitution test).2
The acceptability of the indicative in the examined sentences seems to be related
to the type of hypotheticality expressed by the sentence, i.e., whether a potential or
an irreal event was expressed. Irreal events can be usually expressed only by the past
or present conditional. For instance, in sentences (1) and (2), which both render irreal
events, the present and past conditional verb forms are used, respectively; the conditional cannot be substituted for indicative verb forms here due to the given contexts, cf.
in the sentence (1’) the indicative is unacceptable due to the fact that Joseph Roth is not
alive any more, in (2’) since they (the photos) were not reminded.
On the contrary, when expressing potentiality, the present conditional can be often
replaced by the indicative; see examples (3) and (3’), (4) and (4’). This substitutability
is connected, in our opinion, with the semantic closeness of the potentiality of the
conditional and the future meaning of the indicative (cf. [17]).
2
In [7], the semantics of the past conditional was tested in a similar way.
324
Magda Ševčíková
(1) Svatý pijan Joseph Roth by dnes oslavil rovnou stovku. ‘The saint drunkard Joseph
Roth would celebrate his 100th birthday today.’ (PDT 2.0)
(1’) # Svatý pijan Joseph Roth dnes oslaví rovnou stovku. ‘# The saint drunkard Joseph
Roth celebrates his 100th birthday today.’3
(2) ... připomínat si je by bylo bývalo bolestivé. ‘... to remind oneself of them would
have been painful.’ (SYN2005)
(2’) # ... připomínat si je bylo bolestivé. ‘# ... to remind oneself of them was painful.’
(3) Uhrát tu remízu by bylo úspěchem. ‘To draw the game would be a success.’ (PDT 2.0)
(3’) Uhrát tu remízu bude úspěchem. ‘To draw the game will be a success.’
(4) Podle A. Röschové (ODS) by pro Kozlův návrh byla asi polovina klubu ODS, ale
osobně by byla spíš proti. ‘According to A. Röschová (ODS), approximately a half of
the club of ODS would vote for Kozel’s proposal but she personally would vote against
it.’ (PDT 2.0).
(4’) Podle A. Röschové (ODS) bude pro Kozlův návrh asi polovina klubu ODS, ale
osobně bude spíš proti. ‘According to A. Röschová (ODS), approximately a half of the
club of ODS will vote for Kozel’s proposal but she personally will vote against it.’
3.3
Conditional vs. indicative in selected contexts
However, in the corpus data, several sentences expressing potential as well as irreal
events occurred which violated our hypothesis of the relation of the hypotheticality
type and the acceptability of the indicative. On the one hand, the indicative could be
used instead of the conditional in some irreal contexts. On the other hand, the indicative
was not acceptable in several sentences which expressed potential events. These two
cases could be characterized as follows:
– An irreal event could be expressed both by the (present or past) conditional and the
indicative, for instance, in governing clauses of the examined conditional complex
sentences. In our opinion, the indicative comes into consideration in these sentences
since the irreality is signalized explicitly by another means besides the verbal mood,
e.g., by the conditional clause; see the sentences (5) and (5’).
– Concerning the indicative in potential contexts, the loss of the hypotheticality which
is connected with the examined substitution plays an indispensable role when judging the acceptability of the indicative instead of the conditional. The conditional
mood cannot be replaced by the indicative in such sentences which, according to
our interpretation, express potential events whose realization the speaker does not
accept (cf. the example (6) vs. the sentence (6’), which cannot be accepted in the
given context).
3
The hash mark # indicates that the given sentence (with the indicative) in comparison to the
original sentence (with the conditional) is unacceptable since it does not express a hypothetical
event without regard to the fact that it is mostly a grammatical Czech sentence.
The Meaning of the Conditional in PDT 2.0
325
(5) ... kdyby dnes přišel Thomas Alva Edison do české banky se žádostí o úvěr na výrobu
jakýchsi “žárovek, které změní svět”, skončil by s nepořízenou ... ‘... if Thomas Alva
Edison would come to a Czech bank with an appeal for a credit for production of a
kind of “light bulbs which change the world” in these days, he would go away emptyhanded ...’ (PDT 2.0)
(5’) ... kdyby dnes přišel Thomas Alva Edison do české banky ..., skončí s nepoří-zenou
... ‘... if Thomas Alva Edison would come to a Czech bank ... in these days, he goes
away empty-handed ...’
(6) V každém případě by v Polsku vypukla vážná politická krize. ‘In all cases a serious
political crisis would break out in Poland.’ (PDT 2.0)
(6’) # V každém případě v Polsku vypukne vážná politická krize. ‘# In all cases a serious
political crisis breaks out in Poland.’
To resume the results of the substitution test, the conditional cannot be replaced by
the indicative in most sentences expressing irreal events and in such sentences rendering potential events whose realization is not expected by the speaker. These cases are
considered to be an evidence that the conditional is an irreplaceable means in contemporary Czech texts and has to be included in the representation of the meaning of the
sentence. On the contrary, the conditional can be substituted for the indicative in many
sentences expressing potential events but also in some sentences with irreal meaning.
The relatively free substitutability of the two moods in potential contexts is possible
due to the close relation between the meanings of potentiality and future. In sentences
which express irreal events, the conditional can be substituted for the indicative only
if the irreality is clearly expressed by some other means; these cases do not deny the
semantic relevance of the conditional.
4
4.1
Modality in PDT 2.0 and some other treebanks
Description of modality in PDT 2.0 and FGD
In theoretical works based on FGD, neither verbal mood nor other modal means have
been studied yet in more detail in spite of the importance of these means in constituting
sentence semantics. In PDT 2.0, a special grammateme verbmod was defined. However,
since the values of this grammateme directly correspond to the morphological moods
occurring in the sentence, the grammateme is not to be considered as a semantic counterpart of the morphological category of mood but rather as a provisional, “technical”
solution which requires further investigation.
Within the broad area of modality, main attention has been paid to modal verbs in
FGD (see esp. [12]). On the basis of syntactic and semantic criteria, a group of so called
proper modal verbs were identified; e.g. muset ‘must / have to’, moct ‘can’. Meanings of
these verbs are considered as modal features “added” to the meaning of an autosemantic
verb. In the tectogrammatical tree, meanings of proper modal verbs are represented by a
grammateme belonging to the node representing the autosemantic verb in question (i.e.,
similarly to meanings expressed by the morphological category of tense etc.). Within
PDT 2.0, the modality expressed by modal verbs was captured in the same way as
in FGD.
326
4.2
Magda Ševčíková
Modality in some other treebanks
Also other treebanks, if paying any attention to modality at all, confined to modal verbs
whereas semantics of verbal moods is omitted (the mood is usually reflected just within
the morphological annotation). Semantics of modal verbs is described, e.g., in Proposition Bank, which is a corpus annotated with semantic roles, or in the annotated trebank
of Bulgarian texts BulTreeBank. In Proposition Bank, a special semantic role MOD for
modal verbs was included (cf. [11]). In BulTreeBank, deontic and epistemic reading of
modal verbs was distinguished (see [9]). Also [3], [13], [4] or [15] were concerned with
(semi-)automatic detection of deontic and epistemic usage of modal verbs, especially
in scientific articles.
5
Representing the primary function of the conditional mood at
the tectogrammatical layer of PDT 2.0
5.1
Grammateme of factual modality
As sketched in Section 3, the conditional in its primary function is a semantically
relevant means: by the use of the conditional an event is presented as hypothetical.
In this function, the conditional is opposed to the indicative, by means of which events
are presented as given. The semantic opposition which is expressed by the conditional
and the indicative is thus to be captured when representing the meaning of the sentence,
i.e., within the tectogrammatical tree in PDT 2.0.
Since verbal mood is a morphological category (similarly to that of tense etc.),
we propose to capture the semantic opposition expressed by the conditional and the
indicative by a grammateme which belongs to a verbal node of the given tectogrammatical tree. We introduce a new grammateme of factual modality (factmod). Besides
the opposition of hypotheticality and assertion also the difference between two types of
hypothetical meanings – the potential and the irreal ones – has to be taken into consideration. Three values of the grammateme factmod are therefore proposed: potential for
potential events, irreal for irreal events and asserted for given events.4
5.2
Annotation rules for assigning the grammateme of factual modality
When describing the assignment of the values of the proposed grammateme, we proceed from the surface structure of the sentence in a way similar to that in which a
tectogrammatical representation is assigned to a sentence. A tectogrammatical node
which represents an indicative verb form (the primary function of the indicative and the
conditional is taken into consideration here and in the sequel) is typically assigned with
the value asserted of the grammateme factmod. When a verb form of past conditional
occurs, the value irreal is filled in the grammateme factmod at the tectogrammatical
node representing this verb form.
4
The solution to represent both oppositions by a single grammateme was preferred to the possibility to introduce two grammatemes just for the economy of the former solution.
The Meaning of the Conditional in PDT 2.0
verbal mood
past conditional
present conditional expressing
potentiality
present conditional expressing
irreality
present conditional expressing
potentiality or irreality (unresolved)
indicative expressing assertion
indicative instead of the conditional
with the meaning of potentiality
indicative instead of the conditional
with the meaning of irreality
327
grammateme factmod grammateme tense
irreal
–
potential
–
irreal
–
potential|irreal
–
asserted
asserted
ant / sim / post
ant / sim / post
irreal
ant / sim / post
Table 1. Rules for assigning the new grammateme factmod. In the 1st column, the mood of the
verb form is given. According to the meaning of this form, the appropriate value of the factmod
is chosen (the 2nd column). In the 3rd column, it is indicated whether one of the values of the
grammateme tense is to be chosen (if not, – is filled in).
Concerning the present conditional, one of the values of the grammateme factmod
is to be chosen by the annotator on the basis of context, knowledge of situation etc.: if
this verb form expresses an irreal event, the value irreal is assigned to the corresponding
node; if a potential event is concerned, the value potential is the right one; in case the
annotator is not able to decide between the two interpretations, both values are to be
filled in (potential|irreal). As for the expressions of irreality, the past conditional and
the present conditional are thus considered as synonymous means: for sentences which
differ just in the conditional form used, identical tectogrammatical representations are
supposed.
Besides these basic cases, we also deal with tectogrammatical representation of
the sentences in which an indicative verb form is used instead of the conditional one
(cf. the results of the substitution test in Section 3). Examples such as the governing
clause in the sentence (7), in which the conditional can be substituted for more than one
indicative form (cf. (7’)), indicate that these sentence pairs are semantically close to
each other, though not synonymous. Therefore, they have to be represented differently
at the tectogrammatical layer. Rules concerning the assignment of the indicative which
occurs instead of the conditional are specified in the next subsection.
(7) ... jestliže bychom se rozhodli stát jadernou zemí, musela by následovat celá řada
kroků ... ‘... if we decide to become a nuclear country, a whole series of steps would
have to follow ...’ (PDT 2.0)
(7’) ... jestliže bychom se rozhodli stát jadernou zemí, musí následovat / bude muset
následovat celá řada kroků ... ‘... if we decide to become a nuclear country, a whole
series of steps has to follow / will have to follow ...’
328
5.3
Magda Ševčíková
Annotation of the indicative in originally conditional contexts
We propose to represent the indicative in sentences rendering irreal events in the same
way as the past conditional (or the present conditional with irreal meaning), i.e., by a
tectogrammatical node with the value irreal in the grammateme factmod; additionally,
the node is assigned with the grammateme tense, in which the temporal characteristic
of the indicative form is captured. Unlike the annotation of the indicative in sentences
expressing irreal events, the value asserted is chosen in the grammateme factmod and an
appropriate value of the grammateme tense is assigned if the indicative is used instead
of the present conditional with a potential meaning.
The reason for a different annotation of the indicative in irreal and potential contexts is the disparity of conditions under which the indicative can be used instead of
the conditional in these two context types. In sentences rendering irreal events, the
conditional can be substituted for the indicative only under relatively strong conditions
(if the irreality is clearly expressed by some other means) while the conditional and
the indicative are mostly freely interchangeable in potential contexts. As mentioned
above, this interchangeability is related to the semantic closeness of the potentiality of
the conditional and the future meaning of the indicative. If we decided to represent the
indicative similarly to the conditional in sentences expressing potential events (by the
value potential in the grammateme factmod in combination with a value of the grammateme tense), it would mean at the same time that we have to resolve at each future
indicative form whether it expresses a future or a potential event, which we consider as
impossible in many cases. Assignment rules for the grammateme factmod are resumed
in Table 1. Examples of verb forms assigned with the values of the grammateme factmod
(and grammateme tense if necessary) are given in Table 2.
ex.
nr.
(1)
(2)
(3)
(3’)
(4)
verb form
gramm. gramm. ex.
factmod tense
nr.
by oslavil
irreal
–
(5)
by bylo bývalo irreal
–
(5’)
by bylo
potential –
(6)
bude
asserted post
(7)
by byla
potential –
(7’)
by byla
potential –
(4’) bude
asserted post
bude
asserted post
verb form
gramm. gramm.
factmod tense
skončil by
skončí
by vypukla
musela by následovat
musí následovat
bude muset následovat
irreal
irreal
potential
potential
asserted
asserted
–
post
–
–
sim
post
Table 2. Verb forms from the example sentences (1) to (7’) and corresponding values of the
grammatemes factmod and tense which are to be assigned to the tectogrammatical nodes representing the verb forms in question. Only verb forms which were written in bold in the examples
are included in the table. Verb forms from the sentences marked with a hash mark (i.e., (1’), (2’),
and (6’)) were not assigned.
The Meaning of the Conditional in PDT 2.0
6
329
Conclusions
The present paper focused on the Czech conditional mood when expressing hypothetical events (we talked about the primary function of this mood). It was illustrated with
several examples that the conditional is a semantically relevant means in contemporary
Czech although it could be substituted for the indicative in some of the studied contexts.
The conditional in its primary function should thus be included in the tectogrammatical
representation of PDT 2.0.
For this purpose, the grammateme factmod was proposed. Three values of this
grammateme were suggested by means of which the difference between hypothetical
and asserted events as well as the difference between two types of hypothetical events
(the potential and irreal ones) are captured. Rules for manual assignment of the values
were further described in the paper.
In the near future, rules for (at least) semi-automatic assignment of this grammateme
have to be specified. Also tectogrammatical representation of other functions of the
conditional and further modal means is to be elaborated.
Acknowledgements
I would like to thank Professor Jarmila Panevová for extensive discussion on the presented topic. I thank also Professor Eva Hajičová and Zdeněk Žabokrtský for valuable
comments on the draft of this paper. The work reported on in this paper was supported
by the project 1ET101120503.
References
[1] Bybee, J. L. (1985). Morphology: A Study of the Relation between Meaning and
Form. Benjamins, Philadelphia.
[2] Daneš, F., Hlavsa, Z., Grepl, M. a kol. (1987). Mluvnice češtiny 3. Academia,
Praha.
[3] Danilava, S. and Schommer, C. A Semi-Automatic Framework to Discover Epistemic Modalities in Scientific Articles.
[4] Gabrielatos, C. and McEnery, T. (2005). Epistemic Modality in MA Dissertations.
Lengua y Sociedad: Ivestigaciones recientes en lingüística aplicada. Lingüística
y Filología, 61:311–331.
[5] Grepl, M. a Karlík, P. (1998). Skladba češtiny. Votobia, Olomouc.
[6] Hajič, J., Panevová, J., Hajičová, E., Sgall, P., Pajas, P., Štěpánek, J., Havelka,
J., Mikulová, M., Žabokrtský, Z., and Ševčíková Razímová, M. (2006). Prague
Dependency Treebank 2.0. Linguistic Data Consortium, Philadelphia.
[7] Karlík, P. (1983). K významu kondicionálu minulého. Slovo a slovesnost, 44:12–
21.
[8] Komárek, M., Kořenský, J., Petr, J., Veselková, J. a kol. (1986). Mluvnice češtiny
2. Academia, Praha.
330
Magda Ševčíková
[9] Osenova, P. and Simov, K. (2003). The Bulgarian HPSG Treebank: Specialization
of the Annotation Scheme. In Proceedings of the Second Workshop on Treebanks
and Linguistic Theories (TLT 2003), pages 129–140, Vaxjo, Sweden. Vaxjo University Press.
[10] Palmer, F. R. (2001). Mood and Modality. Cambridge University Press, Cambridge.
[11] Palmer, M., Gildea, D., and Kingsbury, P. (2005). The Proposition Bank: An
Annotated Corpus of Semantic Roles. Computational Linguistics, 31:71–106.
[12] Panevová, J., Benešová, E. a Sgall, P. (1971). Čas a modalita v češtině. Univerzita
Karlova, Praha.
[13] Piqué, J., Posteguillo, S., and Andreu-Besó, J. V. (2001). A Pragmatic Analysis
Framework for the Description of Modality Usage in Academic English Contexts.
Estudios de Lingüística Inglesa Aplicada, 2:213–224.
[14] Sgall, P., Hajičová, E., and Panevová, J. (1986). The Meaning of the Sentence in
Its Semantic and Pragmatic Aspects. D. Reidel Publishing Company, Dordrecht.
[15] Thompson, P., Venturi, G., McNaught, J., Montemagni, S., and Ananiadou, S.
(2008). Categorising Modality in Biomedical Texts. In Proceedings of the LREC
2008 Workshop on Building and Evaluating Resources for Biomedical Text Mining, pages 31–34, Marrakech, Morocco. ELRA.
[16] Ševčíková, M. (2009). Funkce kondicionálu z hlediska významové roviny. Disertační práce. MFF UK, Praha.
[17] Šmilauer, V. (1966). Novočeská skladba. SPN, Praha.
The Creation of the Morphological Ambiguity
Depository in Ukrainian
Olga Shypnivska1 and Sergij Starykov2
National Taras Shevchenko University of Kyiv1 , Ukraine
Ukrainian Lingua-Information Fund of National Academy of Sciences of Ukrainian2, Ukraine
Abstract. Our paper deals with the morphological homonyms as a type of word
sense disambiguation. The study provides a number of linguistic databases of
morphological homonyms as a research environment. A brief description of a
formal classification of the Ukrainian morphological homonyms according to
different types and patterns, frequency dependence of these types from number
of components and grammar properties are presented. The principles and
structure of the electronic dictionary of morphological homonyms presented
both in the dictionary of the word forms and in the stylistic differentiated texts
corpus are described.
1 The morphological homonyms as a type of word sense
disambiguation in Ukrainian. Some problems and proposed
solutions
Ambiguity, the phenomenon that a word has more than one sense, poses difficulties
for many current natural language processing (NLP) systems. The potential for
ambiguous readings is completely unnoticed in normal text and flowing conversation.
Words may be polysemous in principle, but in actual text with contextual knowledge
and topical associations there is very little real ambiguity – to a person. In the
computational linguistics word sense disambiguation is considered as the problem of
computationally determining which “sense” of a word is activated by the use of the
word in a particular context.
Ukrainian as many Slavic is very flexible. Ambiguity at the morphological level is
very widespread. Analysis of the morphological tagging texts shows that more than
70 % word forms are morphological homonyms.
Consider the following sentence taken from the corpus, each containing 29 word
forms:
За <Ь0Н0ПВПРПТ> результатами <ЙЮ> дослідження <ЛИЛРЛВЛАЛУ>
видно <X0>, що <Ь0СПМQМU> серед <КЕПР> так <СПССЬ0Н0КЕ> званих
<АЕАУАЯ> простих <АЕАУАЯ> громадян <ЙЕЙУ> явно <Н0>
проглядається
<ГЯ>
бажання
<ЛИЛРЛВЛКЛАЛУЛШ>
мати
<ГФЙАЙУЙШКИКК> "одного <ЧРЧВЧЧОРОВОЧ> господаря <ЙРЙВ> в
<ПВПППР> домі <ЙП>", який <ОИОВ> несе <ГЯ> всю <ОЛ> відповідальність
<КИКВ> і <СПССN0Ь0> за <Ь0Н0ПВПРПТ> внутрішню <АЛ>, і
<СПССN0Ь0> за <Ь0Н0ПВПРПТ> зовнішню <АЛ> політику <ЙДЙПЙККВ>.
332
Olga Shypnivska and Sergij Starykov
The application of the morphological parser indicates that 20 word forms are
morphologically ambiguous. For example, the word form дослідження means the
nominative, the genitive, the accusative cases of the neuter noun in singular and the
nominative, the accusative cases of the neuter noun in plural. The word form мати
denotes two grammatical classes – the verb and the noun. Мати we can use as
infinitive of the verb. Мати as a noun has different senses: the nominative, the
vocative of the feminine noun in singular and the nominative, the accusative, the
vocative of the masculine noun in plural. (The tagset consists of two-letter symbols
that correspond word forms with a common set of morphological values. It is
described in [9].
Current researches in Slavic, reported in the literature, provide some descriptions
for the morphological ambiguity [3; 4]. There are a lot of works which represent
attitudes and methods of its resolving [4; 5; 6; 7; 10]. But sometimes (in articles
concerning inflexible languages) it argued that homographs are generally much easier
to disambiguate than polysemous words [2]. Little researches have been devoted to
Ukrainian [7]. These studies use rule-based approaches.
Each word relies on different kinds of knowledge for disambiguation. Senses of
ambiguity words can be related in different ways. Therefore it is necessary to define
all possible types of relations between morphologically ambiguous word forms. The
results can be used to identify all possible patterns of disambiguation which expect
valid rules their resolving.
The profusion of different types of homonyms induces us to create of the
depositary of morphological ambiguity in Ukrainian language. In the paper we present
a formal classification of the morphological homonyms according to different types
and patterns. We describe a number of linguistic databases of morphological
homonyms as a research environment and demonstrate the qualitative and quantitative
description of types and patterns of the morphological homonyms in contemporary
Ukrainian. Besides, the electronic dictionary of morphological homonyms is
presented.
2 Principles, structure and the main goal of linguistic databases of
morphological homonyms in Ukrainian
We consider the depositary of morphological ambiguity in Ukrainian as a research
environment. These linguistic databases generally designed:
•
•
•
to represent a list of potential morphological homonyms in Ukrainian both
homonyms belonging to different part-of-speech and homonyms belonging to one
part-of-speech
to show types and patterns of the morphological homonyms and their occurrence
probability in the real texts
to possess all contexts for every pattern.
The Creation of the Morphological Ambiguity Depository in Ukrainian
333
But the central question is how to form a complete list of potential morphological
homonyms. Ukrainian is very flexible and there are a lot of open class words.
Standard dictionary do not contain morphologically ambiguous word forms. There is
a need for new, machine techniques and good collections.
As a base we used the dictionary of the word forms and the morphological parser.
About the Ukrainian grammar dictionary and morphological parser see [9]. The
dictionary has 3 094 174 word forms and 1 554 372 of them are morphologically
ambiguous. From this list we selected 610 612 morphological homonyms.
The main aim of our attitude is to define possible differences among homonyms.
We distinguish two kinds of homonyms: homonyms belonging to different part-ofspeech (for our example – word мати) and homonyms belonging to one part-ofspeech (in our examples – word дослідження). In the course of the research it was
observed that only 2 % (10 215) of the list are homonyms belonging to different partof-speech and 98 % (600 409) are homonyms belonging to one part-of-speech. The
integration of homonyms with the same grammar senses allows us to determine 1 486
patterns from which 820 describe homonyms belonging to different part-of-speech.
Based on the list of patterns according to the part-of-speech meanings, we have
selected 92 types of morphological homonyms (81 of them represent homonyms
belonging to different part-of-speech). Table 1 displays a fragment of these linguistic
databases.
Word form
мов
ніж
поверх
плач
моделі
фізика
Pattern
ГЬКЕСПґЬ
ГЬКЕСПґЬ
ЙВЙИН0ПР
ГЬЙВЙИ
КАКДКПКРКУКШ
ЙВЙРКИ
Type
V-N-Prepos
V-N-Prepos
N-Prepos-Adv
V-N
N
N
Table 1. The linguistic databases of morphological homonyms in Ukrainian
The first column provides morphological homonyms. Columns second and third
provide patterns and types which correspond with them.
Linguistic information represented in this way can be combined in different manners
according to your requirement. For example, we can take homonyms two of the most
frequent patterns with the grammatical meaning of noun and verb. In this pattern verb is
used in imperative and noun has a form of masculine. See Table 2 for examples.
334
Word form
голуб
вихри
бубни
плач
виїзди
твори
Olga Shypnivska and Sergij Starykov
Pattern
ГЬЙВЙИ
ГЬЙАЙУЙШ
ГЬЙАЙУЙШ
ГЬЙВЙИ
ГЬЙАЙУЙШ
ГЬЙАЙУЙШ
Type
V-N
V-N
V-N
V-N
V-N
V-N
Table 2. The linguistic databases of morphological homonyms belonging to different
part-of-speech in Ukrainian (Patterns)
The linguistic databases of morphological homonyms belonging to different partof-speech can give use the most frequent types of homonyms in Ukrainian. Table 3
displays some basic statistics (the second column provides the quantity of homonyms).
Type
Adj-N
V-N
Adj-V
V-Adv
N-Adv
Adj-Adv
Amount
5682
2406
588
389
342
272
Word form
вартові, виборні
коти, лови
бадьорим, білим
значимо, мислимо
візаві, ажур
балакуче, захоплююче
Table 3. The linguistic databases of morphological homonyms belonging to different
part-of-speech in Ukrainian (Types)
Verbs, nouns, adjectives are the most flexible and most open grammar classes in
Ukrainian. For examples, in the Ukrainian grammar there are more then 76 000
nouns. That’s why 6 two-component types (see Table 3) of homonyms with these
part-of-speeches present 95 % of all types and types which have more then two
components coverage only 5 % of homonyms belonging to different part-of-speech.
Studying the data, we observed that near 75 % of homonyms can have similar
components in their meanings. Let us compare, for examples, homonym плач (see
Table 2) it can be as verbal noun and verb with common components of meaning and
homonym мов which can denote as different part-of-speeches (noun and preposition)
and can differ number and gender for noun мова. We distinguished two kinds of
types: complex semantic types and simple formal types.
2.1 Linguistic resources for homonyms functioning investigations in texts and
tools for resolving their ambiguity
To appreciate the quantity of homonyms and their occurrence probability present in
contemporary texts (fiction, journalistic genre, scientific texts), we processed the
3 000 000 word forms added to the database with potential types. This material shows
The Creation of the Morphological Ambiguity Depository in Ukrainian
335
us that 62 % of potentional homonyms belonging to different part-of-speech were
occurred at the texts. The most frequent types include unflexible part-of-speeches:
conjunction, preposition, interjection. These types are not productive in the dictionary
of the word forms and they are not very regular. Though there is clearly a need to
explore the most regular types. As far as verb and noun are the most informative partof-speeches the type Verb-Noun is very regular. For example, this type of homonyms
makes 34 % of all homonyms in journalistic genre. To compare their regularity, types
include unflexible part-of-speeches constitute only 3 %.
Table 4 displays some basic statistics for the most frequent types.
Type
Interj-Conj-Part
Middle occurrence
probability
Type
Middle occurrence
probability
13,83
Type
Middle occurrence
probability
V-N
5,05
Adj-Adv
1,75
Conj-Part
7,89
Adj-N
3,84
N-Pron-Adv
1,28
Interj-Prepos-Part
7,39
Pron-Conj
3,64
N-Pron
1,19
N-Prepos
7,12
N-Adv-Conj-Part
3,48
V-Prep
1,04
Interj-Prepos
5,75
Adv-Prepos-Part
3,07
Adv-Conj-Part
1,02
N-Pron
5,64
Pron-Part
2,35
Pron-Conj-Part
5,19
N-Adv
2,28
Table 4. Middle occurrence probability of the homonyms belonging to different
part-of-speech.
There are a lot of previous works which deal with the disambiguation resolving.
They introduced us different methods such as knowledge-based, unsupervised corpusbased, supervised corpus-based. In general see Eneko Agirre [2]. As a rule they
concern lexical disambiguation in inflexible languages [2]. As for Slavic all studies
use rule-based approaches [5; 6; 7; 8]. But for Ukrainian there is a need finding
attitudes and tools which allow getting results to a high degree of accuracy according
to homonyms qualities and their frequency in our language.
To achieve a compromise between main directions we suggest linguistic databases
of contexts. In our attitude we combine statistics and a method of rules. These
linguistic resources suppose its application as combinations of technics for resolving
morphological disambiguation.
The linguistic databases of contexts consist of two tables. The first one represents
formal grammar contexts for each homonym. The second represents particular formal
grammar contexts for each pattern. These linguistic databases were a two-step
process. The morphological annotated texts served as the base of contexts. In step
one, automatically we define formal grammar contexts for each homonym and
compute them. The length of a context is 7 words in the right and in the left from a
homonym. In step two, we manually determine the particular formal grammar context
for each instance and compute them too. We extract relevant grammar, semantic
features for a given homonym. Also we point the length and the location of these
336
Olga Shypnivska and Sergij Starykov
determining features and the relevant meaning for each instance. In this case
automatically we unite the same grammar contexts. Finally, the linguistic databases of
contexts were integrated with the general linguistic databases of morphological homonyms. Table 5 displays a fragment of the linguistic databases of particular formal
grammar contexts for each pattern.
Pattern
ГЬЙАЙУЙШ
ГЬЙАЙУЙШ
ГЬЙАЙУЙШ
Particular formal
grammar context
(PFGC)
<АААУ>
Relevant
meaning
Noun
PFGC
length in
the left
1
PFGC
length in
the right
0
< ГЭ >
< Ь0 > <ММО>
<КИ>
Noun
Verb
1
1
0
1
Comments
Adjective is
coordinated
with noun
Table 5. The linguistic databases of particular formal grammar context
The results show that more then 96,6 % of morphological disambiguation can be
resolving in this way. It was examined on new texts. Statistical calculations verified
that particular formal grammar context in the left can disambiguate more then 80 % of
all instances. We must admit, when homonyms have similar components in their
meanings ambiguity can be still present as shown by the examples below (see also
table 2):
“Хочеш <ГСГХ> наблизитися <ґФ> до <ЛАПР> ідеалу <ЙДЙР> -<> плати
<ГЬЙАЙУЙШКАКРКУКШ> ”; “Реквієм <ЙИЙВ> над <ПВПТ> двома <ЧЮ>
найвідомішими <АЮ> в <ПВПППР> світі <КДКПЙП> студіями <КЮ> колись
<Н0ГЬ> і <СПССN0Ь0> плач <ЙИЙВГЬ> над <ПВПТ> втратою <КТ>
національного
<АРАВАЧ>
надбання
<ЛИЛРЛВЛКЛАЛУЛШ>
за
<Ь0Н0ПВПРПТ>
повної
<АЗ>
байдужості
<КРКДКП>
держави
<КРКАКУКШ>”).
In the second sentence the formal grammar context length is not enough for this
homonym.
3 The electronic dictionary of morphological homonyms belonging
to different part-of-speech
As an example of the created database application we offer the electronic dictionary
of morphological homonyms belonging to different part-of-speech. For this time we
know 2 dictionaries for Slavic which represent the morphological homonyms [1; 4].
Our dictionary has several advantages. It is machine readable and corpus-based. The
electronic dictionary was automatically generated and can be integrate to different
current natural language processing systems or to other dictionaries. It can be used to
study many aspects of the morphological ambiquity.
The Creation of the Morphological Ambiguity Depository in Ukrainian
337
The main goal of it is to represent morphological homonyms occurred both in the
word forms dictionary and in the stylistic differentiated texts corpus.
It contains:
• a list of the potential and real morphological homonyms belonging to different
part-of-speech;
• all types and patterns correspond with homonyms;
• word lemmas correspond with each homonyms;
• the occurrence probability of the morphological homonyms belonging to different
part-of-speech present in contemporary texts (fiction, journalistic genre, scientific
texts).
As far as the dictionary has machine readable format it can give us a lot of
answers for different queries. For example we can take a list of word lemmas which
can have homonyms. Table 6 and picture 1 show the format of this dictionary.
Homonyms
Pattern
Type
стелю
ГГКВ
V-N
ніж
ГЬЙВЙИСП
V-N-Conj
Word lemma
стелити
стеля
ніж
ніжити
ніж
Table 6. The database of the electronic dictionary of morphological homonyms
4 Further developments
Our goal was to build universal, accessible language databases for Ukrainian in
cooperation with everyone who wants to participate. The presented here language
databases are now available for further researches and for making morphological
disambiguation systems of Ukrainian homonyms. Besides, machine readable format
of the morphological homonyms dictionary allows integrating it to different dictionaries in order to represent all semantic and functioning features of each word.
Finally, obtained results can be used in comparative researches of Slavic.
338
Olga Shypnivska and Sergij Starykov
Fig. 1. The format of the electronic dictionary of
morphological homonyms (fragment)
The Creation of the Morphological Ambiguity Depository in Ukrainian
339
References
[1] D. Buttler editor. (1984), Słownik polskich form homonimicznych, Wrocław –
Warszawa – Kraków – Gdańsk – Łódź, Wyd-wo Polskiej Akademii Nauk.
[2] E. Agirre and Philip Edmonds editor (2007), Word sense disambiguation.
Algorithms and applications http://www.wsdbook.org/
[3] El. Awramiuk (1999), Systemowość polskiej homonimii międzyparadygmatycznej. Białystok: Wyd-wo Uni-tu w Białymstoku.
[4] J. G. Anoshkina (2001), Slovar omonimichnych slovoform russkogo jazyka,
Moskva. http://www.irlras-cfrl.rema.ru/homoforms/
[5] J. V. Zinkina, N. V. Pjatkin, O. А. Nevzorova (2001), Razreshenije funkcionalnoj omonimiji v russkom jazyke na osnove kontekstnych pravil. //
http://www.dialog-21
[6] Ł. Dębowski (2001), Tagovanie i dezambiguacija morfosyntaksyczna.
Przegląd method i oprogramovania, Warszawa.
[7] T. A. Graznuchina, L. G. Bratyscczenko, N. P. Darchuk, V. I. Krytska, T. К.
Puzdyreva, L. V. Orlova (1989), Shlachy unuknennja omonimiji w systemi
automatychnogo morphologichnogo analizu [in:] Movoznawstvo, №5.
[8] T. J. Kobzareva, R. N. Afanaseva, (2001) Universalnyj modul predsintaksicheskogo analiza omonimiji chastej rechi w russkom jazyke na osnove
slovara diagnosticheskich situacij. // http://www.dialog-22.ru
[9] V. A. Shyrokov editor (2005), Korpusna lingvistyka, Кyiv.
[10] M. Rudolf (2004), Metody automatychnej analizy korpusu tekstów polskich,
Warszawa.
Frequency of Words and Their Forms
in Contemporary Slovak Language
Based on the Slovak National Corpus
Mária Šimková and Miroslav Ľos
Ľ. Štúr Institute of Linguistics,
Slovak Academy of Sciences, Bratislava, Slovakia
Abstract. Highest ranking words, word forms and n-grams based on their
absolute frequencies are presented and compared with respect to the size,
time of conception and stylistic focus of the source corpora. Corpus coverage
by the most frequent words is discussed and distribution of words among
morphological categories is analyzed. The dependencies of word use on style
and genre of text are uncovered, utilizing the morphological tagging available
in the Slovak National Corpus.
1 Introduction
The development of Quantitative (Statistical) Linguistics has seen more than 100
years since publication of the first European frequency dictionary (F. W. Kaeding:
Häufigkeitswörterbuch der deutschen Sprache, 1897), and more than forty years since
the first computer-aided compilation of language statistics from an electronic corpus
was presented (H. Kučera & W. N. Francis: Computational Analysis of Present-Day
American English, 1967). Many more languages have since been augmented with a
frequency dictionary and/or statistical characteristic of their own, satisfying demand
coming not only from disciplines such as NLP, (both monolingual and translation)
lexicography, first and second language education, logopeadics, neurology,
psychology etc., but also from language hobbyists interested in various quantitative
properties of particular language units as well as the language itself. The existence of
multiple voluminous corpora and the tools to process them allow for rapid collection
of such statistics, often inspiring further research.
2 Word, word form and n-gram frequencies
J. Mistrík ([1] 1969, [2] 1985) manually analyzed word, word form and language
construct frequencies in Slovak language, based on a text of 1 million proper words
(i. e. without punctuation and any other non-word tokens). Some of his findings
have been confirmed by data from the Slovak National Corpus1, whose latest main
corpus – prim-4.0 – made available in early 2009, counts about 550 million tokens.
For example, the five lemmas (a, v, na, sa, byť) are consistently the top five, or
among the top ten, highest-ranking lemmas across all its specializations, even in the
spoken corpus.
1
http://www.korpus.sk
Frequency in Contemporary Slovak Language Based on the SNC
Corpus
FSS 1969
prim1 2004
prim-4.0 2009
s-hovor 2009
Size
1 mil. w
200 mil. t
550 mil. t
434,676 t
1.
a
byť
byť
byť
2.
byť
v
a
to
3.
sa
a
v
a
4.
v
sa
sa
že
5.
na
na
na
sa
6.
on
ten
to
tak
7.
ten
ktorý
ktorý
ja
8.
že
s
s
v
9.
z
z
že
na
10.
ako
že
z
no
341
Table 1. Ten highest-ranking lemmas in general corpora and in the spoken corpus
Individual style of some authors, or smaller, specific texts, utilizing some
characteristic language constructs (cf. Šimková 2008 [3]), could present statistics
skewed from those obtained from general corpora. We created three corpora of
distinct size and specialization: journalism (one month of the newspaper Smena),
nonfiction (selection of texts regarding belief) and fiction (works of A. Habovštiak).
The top rankings of lemmas from these corpora were essentially consistent with the
general corpora.
Corpus
Smena 1968
belief
Habovštiak
Size
296,917 t
14,697,297 t
1,088,695 t
1.
a
a
sa
2.
v
byť
a
3.
byť
sa
byť
4.
sa
v
on
5.
na
na
na
6.
že
ktorý
v
7.
to
on
že
8.
ktorý
že
aj
9.
s
s
čo
10.
z
to
keď
Table 2. Ten highest-ranking lemmas in specialized corpora
342
Mária Šimková and Miroslav Ľos
Next to the top five lemmas, the corpora usually contain pronouns ten or to,
ktorý (less often used as a conjunction), conjunction že, and prepositions s, z. Fiction
works stand apart from other texts, for example in the higher use of the word on
(he) both in the Habovštiak corpus and the corpus of Mistrík, mainly consisting of
fiction.
Also notable is the higher use of the conjunction and particle aj (too, also, even)
and the diminished conjunctive role of the pronoun ktorý (which), that has likely
been partially replaced by the pronouns čo (that) and even keď (as, when),
suggesting here a fundamental difference in fiction vs. nonfiction authors’ selection
of sentence structures.
The four most frequent words in the corpus prim-4.0 after the verb byť (to be)
are also present in the top ten of both bigrams (a one time, 3× v, 5× sa and 2× na)
and trigrams (3× v, 2× na and sa). These top-tens also contain the most frequent
substantive rok (year), ranking 9th among bigrams and 3rd or 4th (ranked by na
druhej strane if counting ignoring character case) among trigrams. Analysis of the
various texts shows that time units in general (e. g. rok, týždeň, deň, hodina) have
substantial frequencies in all works independent of the work’s style and genre.
Notable among the most frequent trigrams are secondary prepositions (ranking
2nd and 4th, respectively), which is likely a result of the high proportion of journalistic texts in the subcorpus prim-4.0-public, the source of the list of bi- and
trigrams in table 3.
Rank
Bigram
Frequency
Trigram
Frequency
1.
nie je
392,383
že je to
35,800
2.
sa v
315,831
v súvislosti s
35,021
3.
sa na
298,802
v tomto roku
32,553
4.
je to
295,003
v porovnaní s
27,489
5.
že sa
282,920
to nie je
26,740
6.
som sa
277,091
Na druhej strane
26,255
7.
by sa
258,065
na druhej strane
23,464
8.
av
245,519
by som sa
21,807
9.
v roku
235,353
Nie je to
20,509
10.
na to
203,130
že by sa
20,411
Table 3. Ten highest-ranking bigrams and trigrams in prim-4.0-public
Whenever corpora differ in size by multiple orders of magnitude, the statistics
collected on them naturally diverge in several parameters, such as the accumulated
share of top-ranking lemmas (see Table 4).
Frequency in Contemporary Slovak Language Based on the SNC
Ranks
FSS
prim-4.0
Difference
1. – 10.
18.6 %
17.98 %
- 0.62
11. – 20.
6.4 %
5.74 %
- 0.66
1. – 20.
25.0 %
23.72 %
- 1.28
21. – 30.
4.0 %
3.51 %
- 0.49
1. – 30.
29.0 %
27.23 %
- 1.77
1. – 100.
41.5 %
37.81 %
- 3.69
343
Table 4. Corpus coverage by most-frequent words (excluding punctuation)
While according to J. Mistrík’s Frequency Dictionary of Slovak Language
(1969), the top one hundred words cover 41.5 % of the text (when discounting
random or rare words with zero dispersion, this value rises to 56.13 %), this
coverage is almost 4 % lower in prim-4.0. This can be explained by the rising
number of random tokens as a corpus grows in size. It is further supported by
lemmas from prim-4.0 with frequencies 1 or 2, which represent almost 66.5 % of all
lemmas in the corpus, of which an estimated half are numerals, typos, words in
foreign scripts or other symbols.
3 Frequencies of morphological categories
In the corpora based on written text, the first two placed are held by substantives
and verbs (see Tab. 5). The 3rd place is held by prepositions in contemporary texts,
in contrast with pronouns in the corpus of Mistrík, which may signify a higher level
of abstraction in the discourse, perhaps as a consequence of the higher share of
journalistic and professional works in prim-4.0. The ranks of adjectives, conjunctions, numerals and interjections is the same between these corpora; the ranks of
the latter two are also shared with the spoken corpus. The generally low share of
interjections in all corpora is nonetheless much higher in the spoken corpus (0.31 %)
than in prim-4.0 (0.058 %). As was to be expected, verbs and pronouns take the first
two places of the spoken corpus; moreover, the share of conjunctions and particles
is quite higher. The statistics on adverbs shows across all corpora, that their
potentiality (being derivable from both adjectives and substantives) in the language
system is much higher than actual use.
344
Mária Šimková and Miroslav Ľos
Corpus
FSS 1969
prim-4.0 2009
s-hovor 2009
Size
1 mil. w
550 mil. t
434,676 t
1.
substantives
substantives
verbs
2.
verbs
verbs
pronouns
3.
pronouns
prepositions
substantives
4.
adjectives
adjectives
conjunctions
5.
prepositions
pronouns
particles
6.
conjunctions
conjunctions
prepositions
7.
adverbs
particles
adjectives
8.
particles
adverbs
adverbs
9.
numerals
numerals
numerals
10.
interjections
interjections
interjections
Table 5. Frequency of words by morphological category
The highest share of substantives classified by gender belongs to the masculine.
Since the morphological annotation of the SNC distinguishes between animate and
inanimate masculines (see http://korpus.juls.savba.sk/usage/morpho/),
the order of genders thus split is instead as follows: 1 st the feminine, 2nd inanimate
masculine (i), 3rd animate masculine, and 4th the neuter. This order is consistent
between works’ styles and genres, with the exception of professional text (prf),
having the animate masculine gender last, likely due to the generally abstract and
impersonal style of such text (see Fig. 1).
45
40
38,58
38,5
38,37
36,32
36,26
35
30
29,12
30,44
28,96
32,16
27,67
25
20
15
19,27
15,29
10
19,44
18,01
14,45
14,39
img
ski
17,51
16,72
15,79
12,74
5
0
mak
m
i
inf
f
prf
n
Fig. 1. Distribution of substantives by gender (%)
Frequency in Contemporary Slovak Language Based on the SNC
345
The manually morphologically annotated corpus (mak), used as reference in
present analysis due to its negligible error rate, is closest in its gender distribution to
the subcorpora of journalistic and fiction/artistic text (tagged inf and img,
respectively). This reflects the bigger share these two types of texts have in mak (of
the 1 207 939 tokens, 44.3 % are journalistic, 36.7 % fiction, and 19.0 %
professional). The subcorpus ski contains all original text from img created by
Slovak authors. Comparing it to its parent corpus, we see the same share for both
feminine and neuter substantives, but a shift of some 1.3 percentage points in favor
of animate masculines in the Slovak fiction corpus.
The distribution of substantives by number and case (Fig. 2) shows a clear
preference towards singular forms, holding top four ranks in all main four corpora
analyzed. The natural first is Nsg, the common 4 th is Lsg. The second and third
places belong to Gsg and Asg, with not so small differences ranging from 2.5 %
points in mak to more than 7 % points in prf. In fiction, the accusative wins due to
the higher usage of verb-object phrases. The journalistic and professional texts
contain more substantives in genitive, which matches the higher use of genitive
attributes in noun phrases.
s1
s2
s3
s4
s5
s6
s7
p1
p2
p3
p4
p5
p6
p7
24,18
17,98
2,57
15,45
0,16
10,31
6,67
5,43
7,59
0,84
4,95
0,01
2,02
1,85
0
5
10
15
20
s1
s2
s3
s4
s5
s6
s7
p1
p2
p3
p4
p5
p6
p7
25
23,35
14,57
3,14
18,99
0,31
9,78
7,00
5,84
5,63
0,75
6,29
0,01
2,05
2,29
0
5
10
a) mak
s1
s2
s3
s4
s5
s6
s7
p1
p2
p3
p4
p5
p6
p7
24,50
18,95
2,05
13,45
0,02
11,03
5,70
6,25
9,05
0,80
4,59
0,00
2,12
1,49
0
5
10
15
c) inf
15
20
25
b) img
20
25
s1
s2
s3
s4
s5
s6
s7
p1
p2
p3
p4
p5
p6
p7
23,95
20,06
2,17
12,82
0,02
9,77
6,20
6,66
9,22
0,76
4,63
0,00
2,07
1,67
0
5
10
15
20
d) prf
Fig. 2. Distribution of substantives by number and case (%)
25
346
Mária Šimková and Miroslav Ľos
Of plural forms, the genitive enjoys the highest usage (i. e. 5 th place overall),
most visibly again in journalistic and professional texts, followed in these by Npl by
a margin of 3 % points. In fiction, Isg holds the 5th place, then followed by Npl and
Apl. Gpl ranks 8th in fiction. The rarest case is dative, both in singular (9th overall,
10th in pub) and in plural (12th).
Places 9th–11th, containing Dsg, Lpl and Ipl., have only small differences among
each other (0.5 % points in pub and prf to about 1 % point in img). A category of its
own belongs to the vocative, regarded as just a relic in contemporary Slovak
language system due to its near absolute homonymy with nominative. The few
exceptions (e. g. otče, majstre, bratu, priateľu) represent a negligible part of the
corpus, mostly only present in fiction (0.31 % in sg and 0.01 % in pl).
The distribution of prepositions by case (Fig. 3) is naturally distinct from that of
the substantives, as the most frequent nominative completely lacks prepositions,
and both genitive and accusative are often used without any preposition. The first
place is thus taken by locative prepositions, mainly used by journalistic and
professional works. The frequencies of genitive and accusative prepositions, chiefly
used in fiction, are rather close to each other. Fiction works also use instrumental
and dative prepositions more heavily, to express predicate complements.
45
40
41,17
39,99
37,24
35
32,97
33,17
23,94
24,86
22,6
22,2
14,85
14,36
30
25
22,85
20
21,06
15
13,23
10
21,54
21,45
11,31
5,62
5,64
5,4
4,53
mak
img
ski
inf
5
22,44
20,11
11,93
5,53
0
2
3
4
6
prf
7
Fig. 3. Distribution of prepositions by case (%)
From the basic verb forms recognized by the morphological annotation in the
SNC (i. e. B – future of byť, H – gerund, I – infinitive, K – indicative, L –
l-participle, M – imperative), the l-participle (i. e. past tense and conditionals), the
present indicative and the infinitive are the most frequent (Fig. 4). The other three
forms are very specific, with only little applicability. The indicative is the most
frequent in journalistic and professional text (represented by up to one half of all
verbs in the subcorpus), while the l-participle is used the most in fiction (again
nearly 50 % of all verbs).
Frequency in Contemporary Slovak Language Based on the SNC
347
60
50,82
49,37
50
46,12
40
38,29
44,8
35,75
44,73
36,62
39,77
30,35
30
20
15,07
14,65
12,23
11,33
10,89
10
2,88
0
mak
img
B
ski
H
I
K
inf
L
prf
M
Fig. 4. Distribution of verbs by form (%)
The division of verbs by person-agreement is dominated by the third person
singular (Fig. 5), having more than 50 % share across all corpora except Slovak
fiction, where they are just below. In fiction, the second place is held by 1 st person
sg, which drops sharply in other styles towards almost zero in professional text.
Journalistic and professional text has instead established 3rd person pl. as the second.
Considerable is also the share of the indeterminate verb forms (wrt. number and
person), marked under ‘--’ in the chart, where the person is concealed, e. g. due to
omitted subject pronoun (e. g. Prišiel. – ja, ty, on?; Prišli. – my, vy, oni?).
60
50,98
50,64
50
52,51
50,95
49,1
40
30
20
10
15,75
14,19
12,43
4,14
15,25
14,76
6,77
4,52
4,06
19,12
18,06
15,28
5,93
0
mak
img
--
pa
ski
pb
pc
inf
sa
sb
prf
sc
Fig. 5. Distribution of verbs by number and person (%)
348
Mária Šimková and Miroslav Ľos
4 Conclusion
After six years of building the primary database of Slovak National Corpus (made
of texts originating in years 1955 – 2009), its quantitative (550 million tokens), as
well as qualitative (i. e. style and genre structure, lemmatization and morphological
tagging) properties allow for quantitative analysis with significant results. Much
data can also be compared to observations of Slovak language made available 40
years ago, or analogous latest results from other languages.
Linguistic interpretation of frequency-based indicators of select lexical categories
and word forms present in contemporary Slovak language presents one of the first
contributions to this discipline based on the SNC. We have researched absolute
frequencies of presented phenomena based on all written text in the corpus, as well
as their distribution among the three main styles: fiction/artistic, journalistic and
professional. Even the partial results published in present paper signal a marked
tendency of the journalistic and professional styles to merge. We also contrasted
some of the results with data from the early Slovak Spoken Corpus, which is too
small at the moment to have reliable conclusions drawn from.
Detailed information on frequency-based parameters of contemporary Slovak
language will be published in the Frequency dictionary currently in preparation at
the Slovak National Corpus.
References
[1] Mistrík, Jozef: Frekvencia slov v slovenčine. Bratislava: Vydavateľstvo SAV
1969. 728 p.
[2] Mistrík, Jozef: Frekvencia tvarov a konštrukcií v slovenčine. Bratislava: Veda
1985. 320 p.
[3] Šimková, Mária: Jazykové prostriedky vo vybraných dielach Martina Rázusa
(Analýza na báze textov Slovenského národného korpusu). In: Martin Rázus –
politik, spisovateľ a cirkevný činiteľ. Ed. M. Pekník. Bratislava: Veda 2008, p.
238 – 249.
Analysis of the Means Expressing Strong
‘Necessity Not To’ in English and Czech
Based on General and Parallel Corpora
Renata Šimůnková
Technical University of Liberec, Czech Republic
[email protected]
Abstract. The paper presents results of a corpus-based study of basic means
used to express strong deontic ‘necessity not to‘ in English and their
counterparts in Czech. In particular, by means of co-text analysis different
semantic values of these means of expression are delimited. It is further studied
to what extent the choice of particular means is influenced by co-occurring
means at different levels: lexical (collocations), grammatical (colligations),
pragmatic, as well as by the type of discourse.
The matter is first studied within one language and then the structures of
semantic field of ‘necessity not to’ in English and in Czech are compared.
1 Introduction
The initial impulse for this corpus-based study resulted from the work on a
dissertation thesis whose aim was to compare means used to express necessity in
English and Czech based on contemporary fiction and the corresponding published
translations. It has been discovered that the difference in meaning between CANNOT
and MUST NOT is by no means clear and that the generally assumed correlation
between English MUST NOT and Czech NESMĚT, and English CANNOT and
Czech NEMOCI is not always valid. The scope of the study, however, did not provide
enough material to allow the deduction of any general conclusions.
The principal goal of this project is, therefore, to study the basic means used to
express strong ‘necessity not to’ with the use of corpora in both English and Czech in
the following steps:
–
–
–
–
collection of the relevant means from the Czech National Corpus and British
National Corpus
classification of the means according to their different interpretations based
on concrete criteria (e.g. objective/subjective modality, a ban/moral impropriety, etc)
study of the reasons for the choice of particular means based on the study of
co-text (collocations, grammatical, pragmatic and stylistic aspects)
contrastive study of the choice of corresponding means in the two languages
based on the parallel corpus Intercorp
350
Renata Šimůnková
2 Modality
The concept of modality is universal; it is not found only in English or Czech, but in
the majority of, if not in all, languages. According to Palmer (1986: 7) “it is probable
that there are very few languages that do not have some kind of grammatical system
of modality.” It represents an extensive and complex problem and since the fundamental means of its expression differ from language to language, the starting points of
its description can vary between languages.
When trying to define modality in general, the keyword is relation: the relation of
the speaker to the utterance, to the factuality and actualisation, the relation of the
utterance to reality, the relation of the real world to possible worlds. Modality can be
best seen as opposed to factuality – it states the content of the utterance not as a fact
but as a potential fact dependent on certain conditions e.g. the authority and approach
of the speaker (You must take the exam now; you can take the exam no), the reliability
of their judgement (He must have forgotten about the meeting).
Two basic concepts or notions of modality are generally distinguished – those of
possibility and necessity. Each of these two concepts, necessity and possibility, is then
further classified into kinds of modality – deontic, epistemic and often also dynamic.1
One additional feature of modality, which also appears in many works dealing
with modality, is the strength of modality. The strengths are referred to by different
names by different linguists e.g. Huddleston and Pullum (2002: 175) talk about
“weak”,” medium” and “strong”, while Halliday (1991: 182) refers to the same
concepts as “high”, “median” and “low” yet they describe the same matter. On the
basis of strength, the fundamental concepts of obligation and high probability on the
“high” end of the scale and of possibility and permission on the “low” one are
distinguished. Medium modality is somewhere in between, “though intuitively closer
to the strong end than to the weak” (Huddleston and Pullum 2002: 177).
The basic distinction between deontic and epistemic modality is that between
actualisation and factuality. Deontic modality has its illocutionary force and an
utterance including deontic modality has the potential (another important key term
connected with modality) to result in a particular physical act of doing or not doing
something. An utterance including epistemic modality, on the other hand, has the
potential of being/not being true. Halliday (1991: 183) describes deontic modality as
calibrating the area of meaning between Do it! and Don’t do it!, whereas epistemic
modality as calibrating the area of meaning which lies between Yes and No.
An additional kind of modality can be distinguished – dynamic modality. For
dynamic modality the key words are properties (of a situation) or disposition (of a
person). Dynamic modality can be, therefore, seen as more objective than the two
1
Some linguists (F. Kiefer, 1994) distinguish a greater number of different kinds of modality,
but, for our purposes, such a detailed classification is not necessary. (With the pre-set aim in
mind it would even be counterproductive.)
Analysis of the Means Expressing Strong ‘Necessity Not To’ in English and Czech
351
previous kinds. Since, however, a common practice is to distinguish only two basic
kinds of modality such an approach is also adopted in this paper.
The distinction of the area of modality into the two afore-mentioned kinds is not
absolute. In many cases it is a question of gradience rather than of clearly cut
boundaries. The kind of interpretation is usually dependent on context because many
modal expressions can be used with both deontic and epistemic interpretations.
3 Inescapable ‘necessity not to’ in English and Czech – theoretical
foundations
Grammar books do not seem to deal in any big detail with the way in which ‘necessity
not to’ is expressed. It is difficult to find theoretical studies of the differences in the
meaning and use between the individual means. The main difference is the difference
between MUST on the one hand and a set of negative modal means such as
CANNOT, MAY NOT, BE NOT ALLOWED on the other hand, which is described
as the difference between ‘necessity not to do’ and negative ‘permission to do’ = a
ban. The former is felt stronger. This assumption is commented on by Leech (2004),
and Palmer (1979): “Both these statements (with may not and mustn’t) are
prohibitions, but differ in that the second sounds rather more forceful, positively
forbidding instead of negatively withholding permission” (Leech 2004: 95), “But
there is an obvious difference between refusing permission (may not/can’t) and laying
an obligation not to (mustn’t). With the former it is to be assumed that permission is
normally required, while with the latter the speaker takes a positive step in preventing
the action for which permission may not normally be required”(Palmer 1979: 65).
In Czech there are two basic modal verbs used: NESMĚT and NEMOCI, which
since they are both used with external negation express negative permission. The only
formal possibility in Czech to express the meaning of MUSTN’T is MUSET with the
negative infinitive which is rare and restricted. Based on a short comment in Dušková
(1994: 192): “Obvyklý zápor k dispozičnímu may je must not, popř. can’t. Např. you
mustn’t worry nesmíte si dělat starosti, we can’t take the dog into the hotel nemůžeme
(nesmíme) vzít psa do hotelu. One mustn’t be proud. Člověk nesmí být pyšný.“2, one
might arrive at the conclusion that MUSTN’T corresponds to Czech NESMĚT and
CAN’T to Czech NEMOCI. This might be further supported by looking into
dictionaries which either do not deal with the issue at all or suggest the same
correlation, e.g. in Lingea Lexicon NESMĚT = MUST NOT, MAY NOT. The
situation, however, may not be so easy. Grammatically, CANNOT should be close in
meaning to MAY NOT since they both take external negation. One clear distinction
of MAY NOT from the other modal verbs used to express ‘necessity not to’ is the
2
The usual negation to deontic may is must not, or can’t, e.g. you mustn’t worry nesmíte si
dělat starosti, we can’t take the dog into the hotel nemůžeme (nesmíme) vzít psa do hotelu.
One mustn’t be proud. Člověk nesmí být pyšný.
352
Renata Šimůnková
formal nature of MAY NOT. “Zápor may not (v dispozičním významu „nesmět“) se
vyskytuje v úředním jazyce“ (Dušková 1994: 191)3.
As reference materials did not provide any further information, native speakers of
British English were asked for help. The opinions they provided were interesting, and
surprising at the same time. They agreed on the assumption that while there exists a
rather big overlap between MUST NOT and CANNOT, partly depending on a
particular context and intonation, there is a basic difference felt between the two
means, which does not seem primarily to reside in the force of the modal means as
Leech (2004) suggested, but rather in the meaning. While CANNOT expresses either
inability (which is outside the scope of this paper) or something which is officially
(often by law) not allowed/banned, MUST NOT is used to express more personal
involvement of the speaker describing something which is considered morally wrong.
It can be seen as more or less in agreement with Palmer (1971) – (see the paragraph
above). If something is usually not allowed, then one, if they in spite of this want or
need to do it, has to ask for permission. It is, however, not logical to ask for
permission to do something which is considered morally inappropriate. Such things
are usually done either out of ignorance, or because one cannot help it.
Reference material on ‘necessity not to’ in Czech does not offer much information
either. It was, therefore, attempted to deduce the negative meanings of the two modal
verbs (NESMĚT, NEMOCI) from the positive ones. From Benešova’s (1971: 132)
classification of modal verbs it follows that MOCI is used to express possibility in
general – it means in cases when the source of modality are outer circumstances as
well as if the source of modality is a concrete human being and both in the case when
the source of modality is identical with the source of action and when the source of
modality is different from the source of action. SMĚT is used only if the source of
modality is a concrete person and only in cases when the source of modality is
different from the source of action. Based on these findings NESMĚT seems mainly
to expresses things which are not allowed, which are officially banned. NEMOCI, on
the other hand, refers, in addition to inability and impossibility based on outer
circumstances, to things seen as bad, something that a kind of inner control prevents a
person from doing e.g. – to jí nemohu udělat, nemohu ji zradit. According to this
explanation, therefore, the correspondence between MUST NOT, CAN NOT on the
one hand, and NESMĚT, NEMOCI on the other should be completely opposite. An
example from Rowling’s Harry Potter might be used to support this view:
They cannot keep the objects longer than that unless they can prove they are
dangerous. (= it is banned by the Ministry) – Nesmějí si věci z pozůstalosti
ponechávat déle, neprokáží-li …
Since, however, in the same way as in English, also in Czech there is a big overlap
between the two means, there is certain flexibility in their use.
3
Negative may not (in deontic sence: nesmět) is used in official language
Analysis of the Means Expressing Strong ‘Necessity Not To’ in English and Czech
353
The comparison becomes simpler when the discussed modal verbs are used with
past reference because due to formal reasons MUST NOT cannot be used. MUST
does not have the past form and in order to express strong obligation with past
reference other forms must be used, the most common being HAVE TO. HAVE NOT
TO or more precisely HAD NOT TO, however, does not express an obligation not to,
but a possibility not to. Apart from the just mentioned formal reason, there might be a
semantic or logical reason for the non-existence of the past form of MUST with
deontic interpretation. Since MUST, and in the same way probably also MUST NOT,
is strongly subjective and being used when the source of modality is the speaker, it is
highly illogical with past reference (even more so in its negative form). MUST NOT
expresses something seen as wrong by the speaker, so if the speaker is at the same
time the source of the action, the logical result has to be the non-realization of the
action (no space for MUSTN’T here) or the action is accidental and then other means
is preferable, e.g. I did not want, mean, intend… . It is also possible to speak about a
ban in the past and then COULD NOT, BE NOT ALLOWED, BE NOT TO etc. is used.
4 Data collection
‘Necessity not to’ was studied on three corpora: Czech National Corpus (ČNK),
British National Corpus (BNC) and parallel corpus Intercorp. Since the British
National Corpus includes texts mainly from the 1990s, in Czech National Corpus part
SYN2000 was chosen as the corresponding source. The corpus Intercorp, since it is
comparatively smaller, was not reduced what concerns dates of publishing of the
texts, since the issue studied is not likely to be significantly influenced by the time
difference in the order of decades. However only works of fiction were selected and
so was done also in the case of the other two corpora. The individual means studied
were selected on the basis of the above mentioned dissertation. For English the
following means were included: MUST NOT, CANNOT, COULD NOT, MAY NOT,
NOT ALLOWED TO, NOT SUPPOSED TO, BE NOT TO. For Czech NESMĚT and
NEMOCI were studied. In the case of NEMOCI and negative forms of CAN and
MAY which apart from ‘necessity not to‘ express also other meanings (which are in
all the cases considerably more frequent) only cases where these means clearly
expressed ‘strong obligation not to’ were included.
In the cases when the corpora provided a large number of tokens, 250 examples
were processed. In the cases when less than 250 examples were available or
manageable always at least 50 relevant cases were included in the study, with the
exception of MAY NOT for which from 411 cases only 19 were relevant.
The whole issue was first studied for each language separately. The means found
in the corpora were classified into categories on the following criteria: the source of
modality and the source of action (on this basis modality is classified into subjective
or objective) and within each category further two categories were distinguished:
moral impropriety and a ban.
354
Renata Šimůnková
5 Results of the corpus-based study
The results gathered in the way described above are first presented in tables and then
commented on. Table 1 presents data gathered from BNC, Table 2 from CNK.
Modal means
MUST NOT
CANNOT
COULD NOT
MAY NOT
NOT
ALLOWED
BE NOT TO
NOT
SUPPOSED
SUBJECTIVE
Moral impropr.
(in %)
Ban
(in %)
71
57
78
OBJECTIVE
Moral impropr.
(in %)
12
Ban
(in %)
7
17
10
26
22
30
100
12
44
66
48
60
11
8
11
Table 1. Classification of Means from BNC
Modal means
NESMĚT
NEMOCI
SUBJECTIVE
Moral impropr.
(in %)
34
28
Ban
(in %)
OBJECTIVE
Moral impropr.
(in %)
19
41
Ban
(in %)
47
31
Table 2. Classification of Means from ČNK
The data from BNC confirm that MUST NOT expresses mainly subjective
modality, which is generally accepted truth and they also confirm the assumption of
the native speakers discussed above that MUST NOT is predominantly connected
with activities seen as morally inappropriate. Rather surprising are the findings
connected with the modal verb CAN. In both present and past forms it mainly
expresses subjective modality and moral impropriety. The higher percentage of moral
impropriety meanings connected with the past form of CAN can be explained by the
fact that since MUST NOT, which is supposed to be a chief means to express moral
impropriety, does not have a past form and COULD NOT functions here as a
suppletive form.
Another interesting fact which can be inferred from the data is the finding that for
a ban English often uses other means than modal verbs. NOT ALLOWED TO, NOT
SUPPOSED TO and NOT BE TO all predominantly express objective modality and
a ban.
The relative closeness of meanings of MUST NOT and CANNOT/COULD NOT when
expressing ‘necessity not to’ is further confirmed when the right-hand collocates of the two
forms are studied. Among the 30 most frequent right-hand collocates of MUST NOT 16
Analysis of the Means Expressing Strong ‘Necessity Not To’ in English and Czech
355
were also among the most frequent collocates for CAN/COULD NOT. These were verbs
let, refuse, blame, complain, tell, say, get, risk, leave, allow, accept, pretend, disagree,
ignore. All these activities can rather be seen as morally inappropriate since they are not
generally permitted or banned.
The issue was then studied on the parallel corpus Intercorp. English forms served
as the starting forms in the search and their corresponding Czech translations were
excerpted as well. The results are summarized in Table 3.
English
modal
means
SUBJECTIVE
Moral impropr.
(in %)
Must not
Nesmět – 74
no C (means no
explicit modal
means in Czech) – 5
nemoci – 38
nesmět – 6
nelze upřít – 6
nemoci – 60
nesmět – 12
no C – 4
Can not
Could not
Ban
(in %)
OBJECTIVE
Moral
impropr.
(in %)
Ban
(in %)
nesmět – 21
nesmět – 2
nemoci – 1
nesmět – 13
nesmět – 37
nemoci – 3
nesmět – 1
zakázat – 1
nemoci – 9
nesmět – 7
Not
allowed
nesmět – 80
nedovolit – 20
May not
Not
supposed
nemít – 16
nemít – 21
nesmět – 100
nemít – 21
nesmět – 16
Be not to
nemoci – 7
nedovolit – 7
imperrative - 7
nemít – 10
nesmět – 11
no C – 5
nemít – 7
mít zakázáno – 43
nesmět – 29
Table 3. Means from Intercorp
In the distribution of basic interpretations of English means expressing ‘necessity
not to’ the data from the Intercorp correspond to those gained from BNC. As far as
their translations into Czech are concerned, MUST NOT is mainly translated as
NESMĚT. NESMĚT is, however, also the most frequent translation for CAN NOT.
The data from Intercorp show that objective ‘necessity not to’ is in English frequently
expressed by other means than modal verbs. This claim is further supported by an
additional search in which English means used to translate Czech NESMĚT were
excerpted with the following results: imperative construction – 15%, infinitive
construction – 15%, would not – 9%, not want to – 8%, no C – 8%, not permit – 5%.
All corpora also suggest close correspondence between MUST NOT and NEMOCI
356
Renata Šimůnková
since both these modal means are frequently used to express moral impropriety. This
finding is further supported by the study of right-hand collocates of NEMOCI where
the means expressing activities which are unlikely to be banned such as tvrdit, nechat,
zklamat, říci, upírat, přesvědčovat, souhlasit, žádat, kárat etc4 prevail.
Apart from MAY NOT, which is found mainly in formal academic or legal style
and therefore was highly infrequent in the corpora used (only 19 relevant examples in
BNC and in Intercorp it was found only in one text which was an example of
academic style), the other studied means do not seem stylistically marked.
6 Conclusion
Corpora enable to study certain features of language on a large amount of material
which was not possible or extremely time consuming in the past and thus to
distinguish delicate shades of meaning. The study presented here is an attempt at such
a delimitation of different meanings of means used to express ‘necessity not to’. The
generally assumed distinction between MUST NOT and CAN NOT residing in the
fact that MUST NOT mainly expresses strong prohibition was not confirmed. Strong
prohibition as such seems rarely to be expressed at all, the main interpretation of
MUST NOT being subjective moral impropriety. The assumption that CANNOT
unlike MUST NOT often expresses activities which require permission and therefore
can be banned is valid only to a certain extent since the predominant meaning of
CANNOT is also to express moral impropriety, although, in contradiction to MUST
NOT, it expresses objective modality more frequently.
Learners and non-native users of both English and Czech should be aware of the
oversimplification of the correspondence between MUST NOT and NESMĚT on the
one hand and CANNOT and NEMOCI on the other and choose the corresponding
means carefully and from a larger set of the means available.
Acknowledgement
This research has been supported by the Inner Grant Competition of FP TUL, grant
IGS 69/2009.
References
Benešová, E., J. Panevová, and P. Sgall. (1971) Čas a modalita v češtině, Praha:
Universita Karlova
Čermák, F., R. Blatná. (2005) Jak využívat Český národní korpus. Praha: Vivas.
Dušková, L. at al (1994) Mluvnice současné angličtiny na pozadí češtiny, Praha:
Academia
Halliday, M., A., K. (1991) An Introduction to Functional Grammar, London: Arnold
4
Claim, let, dissapoint, say, deny, persuade, agree, ask, reproach
Analysis of the Means Expressing Strong ‘Necessity Not To’ in English and Czech
357
Huddleson, R. and G. K. Pullum (2002) The Cambridge Grammar of the English
Language, Cambridge: Cambridge University Press
Hunston, S. (2002) Corpora in Applied linguistics. Cambridge: Cambridge University
Press
Leech, G. (2004) Meaning and the English Verb, London: Longman
Palmer, F., R. (1979) Modality and the English Verb, London: Longman
Palmer, F., R. (1986) Mood and Modality, Cambridge: Cambridge University Press
Palmer, F., R. (1995) Negation and the modals of possibility and necessity, in
Modality in Grammar and Discourse, edd Bybee J. and S. Fleischman, Amsterdam/Philadelphia: John Benjamins, 454 – 471
Research sources
Czech National Corpus: SYN 2000. Prague, 2000. Available on-line from
http://ucnk.ff.cuni.cz
Czech National Corpus: Intercorp. Prague, 2006. Available on-line from
http://ucnk.ff.cuni.cz
British National Corpus. XML edition. Oxford, 2007.
Diatheses in the Czech Valency Lexicon PDT-Vallex
Zdeňka Urešová and Petr Pajas
Charles University in Prague, UFAL MFF UK, Czech Republic
Abstract. An important design element in all lexicons, whether human-oriented
or designed for computer processing, is the variability of forms in which lexical
units described in the lexicon entries can occur in natural language utterances. If
all such forms and variations were to be listed independently in the lexicon, its
size would be enormous and it would be hard to maintain (every change would
have to be copied to many entries). These problems can even multiply in the case
of lexicons for computerized natural language applications, where entries must
be explicitly and formally described in full detail.
As an inherent part of the Prague Dependency Treebank project ([9]; for its theoretical background, see the work of Sgall et al. [33]) a valency lexicon called
PDT-Vallex ([10], [39], [40]) has been created and is publicly available, with
over 8800 verb senses and their corresponding valency frames, linked fully to the
treebank.
When a particular verb sense is used in a diathetic expression (passive construction, reciprocity, resultative or dispositional modality etc.), the surface expression
of verb complements also changes ([40]). While the basic form “transformations”
are well known, it is less obvious how to describe them for all the modalities,
especially for the purposes of computer processing, where everything must be
explicitly stated. We have found that these transformations can be described by
a set of rules, which then allow to keep only a canonical (i.e., the active-voice)
valency frame in the lexicon entry, and use these rules to obtain surface expression
constraints for all the diatheses covered. This formalization have been used in the
formal checking of the Prague Dependency Treebank project and it is used in
other current projects as well.
1
Valency
Before we will concentrate on diatheses in the PDT-Vallex, let us make a little digression into the very notion of valency and diathesis.
1.1
Valency in general
This introductory section reviews some very basic facts about valency. Most writers
on the subject cite Tesnière [38] as the one responsible for introducing the term of
valency into modern linguistics. Tesnière uses the term valency for syntactic analysis of
a sentence, so it was linked also to dependency. Active valency and passive valency are
occasionally distinguished in literature ([22]).1
1
In this article, whenever we simply refer to valency, we mean “active” valency.
Diatheses in PDT-Vallex
359
After the first presentation of valency by Tesnière, the study of valency was taken up
by many scholars, with a wealth of material now available. Since individual authors see
valency from different perspectives, so far no generally accepted definition of valency
exists (Storrer [36]). Generally, valency is understood as a specific ability of certain
lexical units - primarily of verbs - to open free slots for filling in by other lexical units.
By filling these slots a sentence structure is being built. Valency is seen as both syntactic,
semantic, or some combination of them.
The valency terminology is also inconsistent; terms like valence, subcategorization,
intention (in [30]), government, government pattern ([20]), complex sentence pattern
([4]), argument structure ([26]), stereotypical syntagmatic patterns ([32]) etc. emerge.
Naturally, these terms not always denote exactly the same linguistic phenomenon. For
a detailed survey see [40].
1.2
Valency in the Functional Generative Description
Among theories combining the syntactic and semantic approach, the valency theory
developed within the framework of the Functional Generative Description (FGD) is
found. (see e.g. [28], [29] and [16]). It uses syntactic as well as semantic criteria to
identify verbal complementations. In this theory, it is assumed that potentially every
(semantic) verb, noun, adjective and adverb (i.e. every complex node) has subcategorization requirements, expressed by its valency frame. Valency modifications include
all kinds of elements (dependency relations) that can modify a particular lexical unit.
For example, in the sentence Jitka mu daruje knihu (lit. Jitka gives him a book)
the verb darovat (lit. to give) opens a slot for a subject in nominative, i.e. for an agent
(Jitka), then a slot for a dative object, i.e. for an addressee of giving (mu, lit. him), and
lastly a slot for an accusative object, i. e. for an object which is being given (knihu, lit.
a book).
Since the FGD does not work with the notion of “semantic roles” as known from
some of the literature (such as a “runner”, “giver”, “object-given”, see e.g. [15]), the appropriate lexical unit (here the verb) therefore determines—besides the morphological
requirements on arguments—also their semantic properties.
In FGD, we work with TECTOGRAMMATICAL REPRESENTATION of sentences,
which reflect their underlying syntax and certain types of semantic attributes. In this
formalism, the central position in a sentence (or clause) is occupied by a (typically)
finite verb.2 .
In order to sort out the behavior of all word modifications and in order to describe
their character we define the following main basic principles in our valency concept:
– a valency frame is assigned to each verb sense separately,3
2
3
Also some nouns, adjective and adverbs valency frames are recorded in the PDT-Vallex, but
we don’t discuss them in this contribution.
Verb senses are defined rather coarsely, as opposed to some other approaches, such as the
famous WordNet resource. However, it is not excluded that two clearly distinct senses carry
identical valency frames. In other words, senses are not forced to be merged just because their
valency frames are the same.
360
Zdeňka Urešová and Petr Pajas
– criterion for distinguishing inner participants (arguments) and free modifications
(adjuncts),
– criterion for distinguishing obligatory and optional modifications, and
– the concept of “argument shifting”.
According to the type of dependency, any modification can be classified as either
an INNER PARTICIPANT (that is, an argument) or as a FREE MODIFICATION (which
is close to what is known as an adjunct). A given modification of a particular lexical
unit may be—with respect to its particular governing word—either OBLIGATORY (that
is, obligatorily present in the deep, tectogrammatical structure) or OPTIONAL (that is,
not necessarily present). For the obligatory vs. optional distinction, the DIALOG TEST,
described later, is used.
The distinction between inner participants and free modifications is not verb-specific:
if a dependency type is an inner participant, then it is considered an inner participant for
all verbs which it possibly modifies. We have determined that there are five such types
of arguments: actor (ACT), patient(PAT), addressee (ADDR), effect (EFF), and origin
(ORIG). These arguments have also the additional property that they can appear only
once in a given clause headed by the verb (in the particular verb sense) to which they
belong.
Among the 70 complementation types used in the Prague Dependency Treebank, we
identify 36 verb free modification types (adjuncts): adjuncts expressing semantic time
relations: TFHL, THL, THO , TFRWH, TOWH, TPAR, TSIN, TTILL a TWHEN; adjuncts
for local semantic relations: DIR1, DIR2, DIR3 a LOC; adjuncts for causative relations:
ACMP AIM, CAUS, CNCS, COND a INTT; adjuncts for means relations: CPR, CRIT,
DIFF, EXT, MANN, MEANS, REG, RESL a RESTR; modal adjuncts: ATT, INTF a MOD;
semantically different adjuncts: BEN, CONTRD, HER a SUBS, and finally, adjuncts
with double semantic dependency (verb and another verb agrument): COMPL.
While arguments can modify just a relatively closed class of verbs, every adjunct
can modify (in principle) any verb. That is also where their name (free modification)
comes from; moreover, they can be repeated within a given clause several times.
For distinguishing among the five inner participants we use syntactic as well as
semantic criteria. Actor (ACT) is always the first inner participant (something like Arg0
in the PropBank) and Patient (PAT) is always the second inner participant (usually like
Arg1 in the PropBank, see [26]). These two arguments are thus determined more or less
syntactically. Only when a verb has more than two arguments, semantic criteria come
into play. Semantic origin (for example, to make of wood) gets the label ORIG, semantic
addressee (talk to somebody) gets ADDR and semantic result (effect) gets the label EFF
(to split into pieces).
To stress the distinction between the typology of the first two arguments of the verb
and the rest (if any), FGD has adopted the concept of shifting of arguments. According
to this special rule, semantic Effect, semantic Addressee and semantic Origin (which
would normally be labeled by EFF, ADDR and ORIG, respectively) are being shifted
to the Patient position in case the verb has only two arguments. In the sentence Peter
has dug a hole the semantic Effect (a hole) happens to be labeled PAT (as it is in all
sentences headed by the same sense of the verb to dig); similarly, in the sentence The
Diatheses in PDT-Vallex
361
teacher asked the pupil, the semantic Addressee is shifted to the Patient position. This
rule simply helps to keep consistency at the expense of lower semantic “precision”.
Both arguments and adjuncts can be in their relation to a particular word either
obligatory (i.e., obligatorily present at the tectogrammatical level of sentence representation) or optional (i.e. not necessarily present in each sentence where the verb is used).
It should be stressed that this does not concern the surface appearance of such modifications, because they can be elided virtually anywhere; the notion of obligatoriness is
used in the semantic sense. A natural question arises how the obligatoriness can then
be determined, given that surface appearance cannot be used as a criterion: we rely on
a DIALOG TEST [27]. The dialog test is a method based on a question about something
that is supposed to be known to the speaker because it follows from the meaning of
the verb the speaker has used. If the speaker can sensibly answer a hearer’s follow-up
question about a semantically obligatory modification “I don’t know”, then it means
that the given modification is optional. On the contrary, if the answer “I don’t know” is
not possible in the particular point of this dialog-to-be, then the given modification is
considered obligatory. For example, if the verb to leave (in the sense of “departing”) is
used in a sentence John left, the speaker must know from where John left (otherwise,
he or she would—even should—have used another verb). Consequently, “from where”
(DIR1) is an obligatory modification. Conversely, the speaker does not need to know to
where John left—thus, if present, the “to where” (DIR3) modification will always be
optional.
1.3
Valency and the Prague Dependency Treebank
The concept of the valency frames in the Prague Dependency Treebank (PDT) annotation ([21]) corresponds to the valency theory built in the FGD framework described
above.
The work on the valency lexicon enabled the confrontation of the valency theory and
real usage of language. Thus, we can say that PDT-Vallex has been created “bottom-up”;
it was not necessary to make up valency complementation examples for the theoretically
given schemes of valency frames because the lexicon draw upon the real texts from a
real corpus.
Primarily, the PDT-Vallex served for keeping inter-annotator consistency high during the process of manual corpus annotation, most importantly for functor assignment
to verbal complementations. After the tectogrammatical annotation process has ended,
the lexicon served also for rigorous, automatic cross-checking of the annotated PDT
data against this newly built lexicon.
The PDT-Vallex contains only those words (verbs, nouns, adjectives and adverbs)
and their senses which occurred in the annotated data. The lexicon contains 10039 different words: 5510 verbs, 3727 nouns, and a small number of adjectives and adverbs.
The total number of valency frames is 14979, out of which there are 8810 valency
frames for verbs.
The valency modifications are described in the valency frame of the particular verb.
Arguments (inner participants) are always recorded, be they obligatory or optional;
adjuncts (free modifications) are recorded only if determined obligatory.
362
Zdeňka Urešová and Petr Pajas
Apart from the obligatoriness indication we also record the dependency relations
(the FUNCTOR) and the morphemic surface form. Every verb has at least one valency
frame; each frame corresponds to one sense (meaning) of the verb.4
For example, the (English) verb to leave, which has two clearly distinguishable
senses, would have two valency frames in our valency lexicon. The first one would be
used for the sense somebody left something (with an Actor and a Patient as the two
obligatory arguments) and the second one for the sense somebody left from somewhere
(with and Actor and Direction-from as the obligatory arguments for this sense).
Fig. 1. The PDT-Vallex entry for dosáhnout (to reach)
A real PDT-Vallex entry for the verb dosáhnout (to reach) can be seen, formatted
for better readability, in Figure 1. This verb has four different senses in our dictionary.
The first sense to reach something corresponds to the first frame, which contains ACT
in nominative and a PAT in genitive or in the accusative morphemic case. This frame
has been used 272 times in the data. Below the formal description of the valency frame,
three usage examples can be found. Similarly, the other three senses are described using
the same structure.
An example sentence with the verb dosáhnout used as the main verb is in Figure 2.
Obviously, this is an example of the usage of the most frequent sense (first entry as
seen in Figure 1). The Actor (ACT) is the word kurs (price), and the other obligatory
argument is the Patient (PAT)—the word hodnota (value), further modified by the actual
price tag and the currency designation.
In Figure 3, we have schematically depicted how the corpus is linked to the PDTVallex lexicon. Let’s say that in the corpus, we have 3 occurrences of the verb uzavřít
(lit. to close) in 3 sentences. There are two PDT-Vallex entries (valency frames) for
uzavřít. The first two occurrences are linked with the second valency frame with the
basic meaning of to close, which has the usual transitive frame with two arguments:
ACT and PAT. The third occurrence of close is linked with the first valency frame, which
represents the light verb meaning, denoted here with the CPHR functor in its frame. The
4
In rare cases, the description of which is beyond the scope of this article, two or more valency
frames may still be used with the same sense of the verb. This is however not reflected in the
current format of PDT-Vallex. For a suggestion of possible restructuring to allow for (i.a.) such
grouping, see [42].
Diatheses in PDT-Vallex
363
t-ln94206-9-p9s1
root
dosáhnout .enunc
PRED
burza
kurs hodnota
LOC.basic ACT PAT
pražský
RSTR
akcie maximální koruna
APP RSTR
RSTR
Energovod
APP
2130
RSTR
Fig. 2. The (simplified tectogrammatical representation of the) sentence Na pražské burze dosáhl
kurs akcií Energovodu maximální hodnoty 2130 korun. (Lit. On the Prague Stock Exchange
reached the price of shares of Energovod the maximum value of 2130 Czech crowns)
Fig. 3. Links between the corpus and the PDT-Vallex entries
arguments are linked implicitly, and their correctness in form can be determined using
the valency entries and the diathesis transformation rules described in this article.
A detailed description of the surface form of a valency modification as captured in
the valency frames in the PDT-Vallex can be found in Sect. 3.1 of this article, because
this information is relevant for the description of diatheses and their surface transformation rules; for the most detailed information, see the annotation manual ([21]).
2
2.1
Diathesis
Diathesis in general
In contemporary linguistics, the very term DIATHESIS is closely related to other terms,
e.g. alternation ([15]), conversion ([1]), hierarchization ([5]) or genus verbi ([12]). The
364
Zdeňka Urešová and Petr Pajas
phenomenon of diathesis was elaborated in detail in the aforementioned work of F.
Daneš and in other books and articles ([4], [35], [6], among others). Lately, this question
is researched in relation to the valency lexicon [14]. Worldwide, the diathesis issue is
elaborated e.g. by Mel’čuk [19], Chrakovskij [3], Padučeva [24] and [23], Uspenskij
[41] or Babby [2].
The DIATHESIS is understood as a syntactic grammatical category related to verb
voice. It is defined, e.g., as the “relation among the elements of the semantic structure
of the sentence and their corresponding syntactic positions” ([13], p. 522).
A given proposition can appear in the syntactic structure of the utterance either
in its primary (basic, or canonical) diathesis and in a number of secondary (derived)
diatheses. The primary diathesis is defined as the use of the verb in active voice (and in
finite form). More general definition (12) says that the primary diathesis is the one
which co-occurs with the highest number of complements on the surface, with the
subject to the left of the verb.5 Other diatheses, in which the semantic (deep) subject (or
ACT in our case) is not in the subject position in the surface structure, are considered
secondary. In Czech, the secondary diatheses are signaled by specific verb forms (such
as the passive voice) and syntactic structures which force the semantic subject to move
out of the surface subject position.
Specifically, periphrastic or reflexive passive constructions, constructions with the
verb “to have” with verbal passive participle (resultative, similar to perfect tense in English), dispositional modality constructions, or the construction “to get” with a passive
participle (causative) are all examples of secondary diatheses in Czech.
Sometimes, alternations are also classified as diatheses, such as in Kettnerová and
Lopatková [14], where they distinguish grammatical diatheses (roughly, the ones mentioned in the previous paragraph) and semantic diatheses (alternations as the term is
defined elsewhere).
The Czech valency dictionaries, both printed ([37], [17]) and electronic ([18], [25],
[11]), contain (if ever) just canonical forms of the verb complementations used in the
primary diathesis. The only exception is the dictionary of Skoumalová [34], who captures also the explicit complementation forms used in the secondary diathesis. However,
for the natural language processing (as well as for verifying the theoretical language
description) a valency lexicon with a systematical description of all syntactic and morphosyntactic forms is needed; the analysis and synthesis of Czech language, i.e. information about the morphematic realization of particular verb complementations, would
be helpless without this piece of information.
3
Transformation rules for diathesis in PDT-Vallex
In this section, we first describe the means for formalizing form of expression of the
verb and especially its arguments in PDT-Vallex. Then, after describing the general
ideas behind formalizing form transformation occurring in diatheses, we describe the
types of diatheses that we have dealt with in PDT-Vallex in more detail and give some
examples.
5
In Czech, this implies the standard word order, i.e. with the subject not being the focus of the
sentence.
Diatheses in PDT-Vallex
3.1
365
Explicit description of surface form in PDT-Vallex
As it has been said already, the PDT-Vallex entries describe, in a fully formalized way,
the canonical (primary diathesis) expression of the verb and its arguments. However,
with only a few extensions, the same formalism can be used for the explicit description of the necessary form of expression of secondary diatheses (i.e., the result of a
transformation of the canonical surface form to the secondary diathesis one).
In all cases, the description of form should in general describe the predicate and its
arguments as a whole, due to various possible interdependencies among the expression
of form for the verb and its individual arguments ([8])) . However, with only a few
exceptions, the form is independent among the arguments of the verb. Therefore we
decided that the description of form will be associated with the individual arguments,
independently of each other. This is true for both the canonical form (as present in PDTVallex) as well as for the resulting transformed form descriptions for the diatheses, even
though the transformation rules themselves have to consider several or all the arguments
at once.
The part-of-speech and morphosyntactic requirements which characterize the surface form corresponding to the valency slot (argument) in question are denoted by a
short, formally defined string of symbols, separated from the m-lemma (if any) by a
separator (mostly a period, (.)). They appear in the following order: part of speech
requirement, then the morphosyntactic requirements (values) of gender, number, case,
degree of comparison, agreement and negation. If any of these designators is missing,
any value of the given category is allowed (in most cases, that means it is not really
relevant for the relation between the verb and the argument in the corresponding slot).
The first designator (for part of speech) sometimes carries some additional requirement, such (for a verb) to be an infinitive. At the part-of-speech position, clausal restrictions (if the complement is realized as a clause, and not as a noun or other simple
phrase). Lowercase letters are used at this position:
a
d
n
i
v
f
u
j
s
c
adjective
adverb
noun
interjection
verb
verb in infinitive
possessive pronoun or adjective
subordinate conjunction (with a clause it governs)
direct speech (root of subtree)
relative clause (root of subtree)
Gender is denoted by the following four capital letters: F for feminine, M for masculine animate, I for masculine inanimate and N for neuter.
Number is denoted by uppercase S and P with the obvious meanings.
For (morphosyntactic) case, digits are traditionally used for the seven cases in Czech:
1 for nominative, 2 for genitive, 3 for dative, 4 for accusative, 5 for vocative, 6 for
locative and 7 for instrumental. The degree of comparison also uses the digits 1 to 3
366
Zdeňka Urešová and Petr Pajas
for positive, comparative and superlative, respectively, preceded by the symbol @ to
distinguish them from the case markers.
Agreement in gender, number and case with the governing node at the surface layer
is denoted by #.
For negation, we use the tilde character (~).
In addition, any combination of the morphological attribute values that are used at
the morphological and analytical layers of the PDT can be included as a requirement.
Special marking separates them from the above shorthands: a $ symbol and a tag index
in < and > must precede the concrete value, which then should match directly the tag
position at the given index.
The surface form designation might have alternatives (separated by a semicolon),
or even be empty. Empty form designation is allowed only for free modifications, and
it means, that any form that is associated with the particular functor can be used.
Examples of the designation of requirements on the surface realization, roughly
sorted by frequency of appearance in the PDT Vallex dictionary:
1. Case-only requirement: .4
2. Preposition and a particular case: s[.7]
3. Alternative surface expression: preposition (with a particular case), or a case-only
designation: pro[.4];.3
4. A particular subordinate conjunction (alternative of two) governing a verbal clause
on the surface: že[.v];aby[.v]
5. Dependent clause, no conjunction: .v
6. Multiword preposition with genitive: od-1[na-1,rozdíl,.2]
7. Phraseme balit fidlátka (lit. pack one’s belongings, i.e. to leave): fidlátko.P4
It should be also noted that in reality, the form designators described above are
rather short abbreviations for sometimes much more complicated logical expressions;
for example, .1 and .4 are matched also by various prepositional or nominal forms of
numerical expressions not necessarily in nominative or accusative.
In addition, a preposition requiring a single case can be abbreaviated as
<preposition>+<case> (e.g., as in od+2, corresponding to the less readable format
od[.2]). For the purpose of brevity, we similarly introduce (in this paper only) a single
number (for example, 4) to mean a non-prepostitional, direct object in the given case
(normally written as .4).6
3.2
Diatheses and form transformation
Again, the PDT-Vallex dictionary [39], contains only the canonical surface form designation, i.e. the one which describes the form of the verb complements in the primary
diathesis appearance (active voice, finite form). This seemingly causes inconsistency
with the corpus annotation, since the form as required by the form designator in the
valency frame pointed to by the occurrence of the verb sense in the corpus does not
match the actual form of the complements in the corpus if any of the secondary diatheses is used. However, since the change of the form when the verb appears in a secondary
6
This latter abbreviation is not used in the real data, however.
Diatheses in PDT-Vallex
367
diathesis does not depend on the particular verb or verb sense, we can use quite general
“transformation rules”, which can convert the form designation present at the canonical
valency to its “secondary” designation which then should display a perfect match with
the annotated occurrence in the corpus.
This has allowed for a complete verb sense and valency annotation of every occurrence of every verb in the corpus, while still being able to check for the correct relation
between the verbal frame in the PDT-Vallex dictionary and its surface realization in the
corpus even if the verb is not used in the primary diathesis form.
The transformation rules were prepared with annotation consistency checking in
mind. Therefore, they aim at such transformation of the valency frame (using also the
annotation information from the corpus) to arrive at a simple set of checks to either
confirm or reject whether the annotation of the verb and its dependents and its context
is consistent with the requirements found in the valency frame.
Every transformation rule has two parts:
1. condition to be fulfilled at the node being checked, and
2. a set of rewriting rules.
Every rewriting rule has three parts:
1. the type of the rule (replacement, alternative)
2. assertions about the verb frame
3. specification of the necessary changes in the valency frame
While we will not fully dissect all the rules in the following sections describing the
individual diatheses according to the above structure, we will aim to describe all the
main aspects of the transformation. Full details can be found in [40].
3.3
Transformation rules for periphrastic passivization
Only transitive verbs (i.e., verbs with a slot marked PAT in the PDT-Vallex) can appear
in the periphrastic passivization type of diathesis.
The verb itself must be in the form of passive participle, and the actor (ACT) is
moved from the subject position to either an instrumental-case object position (corresponding roughly to the English prepositional phrase with the preposition by) or it is
realized as a prepositional phrase with genitive preposition od (lit. from). Sometimes,
both forms are allowed ((7) as well as (od+2)). This transformation of form always takes
place, regardless of what other complement gets to the subject position.
The (surface) subject can either be missing (zero pronoun form), or in fact any of
the arguments can get to that position:
The painter painted a picture.PAT
→ A picture was painted by a painter.
ADDR The injury slowed down the athlete.ADDR
→ The athlete was slowed down by an injury.
EFF The teacher has read a resume.EFF about him
→ A resume has been read about him by the teacher.
PAT
368
Zdeňka Urešová and Petr Pajas
Periphrastic diathesis, if used with a perfective verb, can be also considered a resultative diathesis, if it describes a completed event. However, the aim of the use has no
reflection on the form changes that the arguments undergo in this diathesis, therefore
we are not making this distinction here.
In periphrastic passivization, which is by far the most frequent of all diatheses, the
following cases are covered by our rules:
1. PAT is moved to the subject position; in this case, we have to further look at the form
of the PAT actually found in the data. The individual forms will be transformed as
follows:
• (4) → (1): nominative case for the “moved” subject
• (f) → (f): infinitive stays as such; verb agreement: 3rd. pers. Sg. Neuter
• (c) → (c): relative clauses do not change; 3rd. pers. Sg. Neuter
• other form of PAT → DELETE (i.e., does not appear on the surface)
• other arguments: forms kept as they are recorded in the canonical frame, except
if EFF(jako+4) (lit. as with accusative) is present in the valency frame together
with PAT(4) → PAT(1), then it is changed to EFF(jako+1).
2. ADDR is moved to the subject position
• (4) → (1): nominative case for the “moved” subject; for other forms of expression in the canonical form, no change.
• other arguments: forms kept as they are recorded in the canonical frame
3. EFF is moved to the subject position
• (4) → (1): nominative case for the “moved” subject; for other forms of expression in the canonical form, no change.
• other arguments: forms kept as they are recorded in the canonical frame (except
when EFF(jako+4) is used, see above at PAT).
Example of transformation:
požádat (to ask)
říci (to say)
přijímat (to hire)
ACT(1) ADDR(4) PAT(o+4,aby)
→ ACT(7) ADDR(1) PAT(o+4,aby)
ACT(1) ADDR(3) PAT(o+6) EFF(4,že)
→ ACT(7) ADDR(3) PAT(o+6) EFF(1,že)
ACT(1) PAT(4) EFF(jako+4)
→ ACT(7) PAT(1) EFF(jako+1)
Example application to a sentence:
Rektor požádal tajemníka o dokumentaci. → Tajemník byl požádán rektorem o
dokumentaci. (lit. The rector asked the secretary for the documentation. → The
secretary was asked by the rector for the documentation.)
Here, the ADDR (the secretary) moves to the subject position and its form changes
from the canonical accusative to nominative (tajemíka → tajemník), whereas the ACT
(the rector) must be then expressed in the instrumental case (rektor → rektorem). The
PAT (documentation) remains in the prepositional phrase form, using the preposition
o (for) and the accusative case (o+4).
Diatheses in PDT-Vallex
369
Univerzita přijímala tyto cizince jako překladatele. → Tito cizinci byli univerzitou
přijímáni jako překladatelé. (lit. The university hired these foreigners as translators. → These foreigners have been hired by the university as translators.)
Both the words cizinci and překladatelé appear in the passive construction in
the nominative case, following the ACT(1) PAT(4) EFF(jako+4) → ACT(7) PAT(1)
EFF(jako+1) rule.
3.4
Transformation rules for reflexive passivization
In Czech, reflexive passivization adds the particle se (lit. itself, but it should be noted
that this word has lost its proper meaning in this construction) to the verb. It can only be
applied to verbs which do not use the same particle in active form (reflexivum tantum),
or inherent reflexives, where the reflexive meaning is lost completely, such as smát se,
lit. to laugh).
In the transformation of form for reflexive passivization, the subject position is taken
by some other argument than ACT, similarly to the periphrastic passivization. However,
the ACT never appears on the surface; it is structurally excluded by the syntactic rules
of Czech. In other words, this diathesis can be used by the speaker only in the case he
does not need to explicitly mention the ACT argument in his utterance, such as in the
case when it is general (“everybody”).
It is quite common that either PAT or ADDR, which would normally be moved to
the subject position, are dropped7 , too. In such a case, a place (LOC, DIR1, ..) or time
(TWHEN, TTILL, ..) expression must be present8 , or at least understood from the context.
This fact, however, does not change the form of transformation rules.
We do not repeat the individual rules for the transformation of form for the PAT,
ADDR, EFF and ORIG arguments, since they are the same as in the periphrastic passivization. However, the following rules must be applied in addition to the periphrastic
passivization ones:
1. the particle se must be added to the verb (as a separate word), with its word order
determined by Czech grammatical rules;
2. the phrase corresponding to ACT must be completely dropped on the surface (i.e.,
at the tectogrammatical representation the ACT must be represented by the #Gen
t-lemma);
3. the tense of the verb remains active (or it remains in the infinitive, as the case may
be);
4. the agreement rules on the surface are also determined by the Czech grammatical
rules (e.g., if subject is not present at all on the surface, the verb must be in 3rd
person singular form, and neuter gender if applicable).
7
8
By “dropping” an ACT (PAT, EFF, ...), we mean that there is no word or phrase in the surface form of the sentence corresponding to the dropped argument. In the tectogrammatical
representation of the sentence, a special t-lema and a special attribute are used to denote this
fact.
Unless the verb itself is in the proper focus of the sentence; for example, as a reply to the
question What do you do with books?, one can say (only) Books are read. without adding
when or where.
370
Zdeňka Urešová and Petr Pajas
5. if PAT is moved to the subject position, the same rules apply as in the periphrastic
passivization, except the ACT must be dropped on the surface.
Example application to a sentence:
Dělníci staví studnu z kamene. → Studna se staví z kamene. (lit. The workers build
the well from stone. → The well [itself] builds from stone.)
The ACT (Dělníci) must be dropped completely, while the PAT (Studna) becomes
the subject (in nominative, as usual).
Univerzita přijímala tyto cizince jako překladatele. → Tito cizinci se přijímali jako
překladatelé. (lit. The university hired these foreigners as translators. → These
foreigners have [themselves] hired as translators.)
The ACT (Univerzita) has been dropped in the reflexive passivization transformation. In addition and identically to the periphrastic passivization diathesis, both the
words cizinci and překladatelé appear in the passive construction in the nominative
case, following the ACT(1) PAT(4) EFF(jako+4) → ACT(7) PAT(1) EFF(jako+1) rule.
Děti o Vánocích zpívají koledy. → Koledy se zpívají o Vánocích. Children at Christmastime sing carols. → Carols [themselves] sing at Christmastime.)
While the PAT undergoes the usual transformation from accusative to nominative,
a time (or location) adverbial is usually kept (or “added”) for the sentence to sound
natural in the reflexive passivization form (for an exception, see the footnote on the
preceding page).
3.5
Transformation rules for resultative diathesis
The resultative diathesis (which is normally expressed by the verbs mít (lit. to have),
dostat (lit. to get) with passive participle of the main verb) is used to move the addressee
to the surface subject position (sometimes also to hide the causativity, or the agent, of
the event). It can only be used with transitive verbs, where addressee must be present
(either as a true addressee (ADDR) with patient (PAT) also present, or shifted to the
patient position). Moreover, the dostat-type of diathesis can be used for only a limited
class of verbs.
The following transformation rules apply (ANY means that the change applies to
whatever the original form of the argument was):
1. ADDR is moved to the subject position (no other argument can be moved to this
position):
• (ANY) → (1): nominative case for the “moved” subject
2. PAT, if it appears on the surface, keeps its form, and if in accusative, forces an
agreement in gender and case of the passive participle of the main verb, since this
argument becomes the complement (Atv, AtvV)9 on the surface. Moreover, if the
9
These are the surface syntax functions. For the description of the formalization of surface
syntax, which is outside the scope of this article, see e.g. [7].
Diatheses in PDT-Vallex
371
gender is feminine in singular, the passive participle must use the special accusative
form -u (the only verbal form which is thought to have a morphosyntactic features
of case). PAT can also be deleted on the surface; the agreement forced on the passive
participle form is then neuter singular.
3. ACT is moved to surface object position with the usual form transformation:
• (1) → (7;od+2)
In the resultative transformation, however, the (od+2) form is much more frequent
than the instrumental case form.
4. Forms of other possible arguments are kept as they are.
Example of transformation (mít, lit. to have):
připravit (to prepare)
ACT(1) ADDR(3) PAT(4)
→ ACT(od+2;7) ADDR(1) ?PAT(4)
Example application to a sentence:
Otec připravil dceři školní tašku. → Dcera měla školní tašku připravenu od otce.
(lit. The father prepared [for] daughter the schoolbag. → Daughter had the schoolbag prepared by the father.)
Example of transformation (dostat, lit. to get):
přidat (to add)
ACT(1) PAT(4;na+6) ADDR(3)
→ ACT(od+2;7) ADDR(1) ?PAT(4;na+6)
Example application to a sentence:
Ředitel přidal na platu jen střednímu managementu. → Jen střední management
dostal od ředitele přidáno na platu. (lit. The director raised in salary only to middle
management. → Only middle management got from the director raised in salary.)
3.6
Dispositional diathesis (dispositional modality)
This type of diathesis is used when the speaker expresses the “modality” (extent, in the
form of and adverbial) of the relation between the verb and its actor, typically when
the actor is an animate object (a human). The adverbial, which must be present on the
surface form once the diathesis is applied, expresses often the degree of difficulty (or
ease) with which the actor can perform the given action or keep themselves in a given
state.
The following transformation rules apply (ANY means that the change applies to
whatever the original form of the argument was):
1. ACT is either elided on the surface (represented as a general actor), or if present, its
form is changed to the dative case:
• (1) → (3): dative case for the actor; often expressed by a pronoun, with no
restrictions on the short/long form of the pronoun (standard grammatical rules
apply).
372
Zdeňka Urešová and Petr Pajas
2. PAT becomes the surface subject, with accusative becoming nominative:
• (4) → (1); other forms unchanged.
3. forms for other arguments are unchanged. This might result in two dative-form
arguments of the verb, which is normally avoided, but it cannot be excluded on
purely grammatical grounds. Often, though, the other arguments are elided on the
surface.
Two other conditions must also be fulfilled:
1. An adverbial expressing the degree of difficulty must be present on the surface, or
implicitly but clearly understood from the context;
2. the particle se must be added to the verb (as a separate word), with its word order
determined by Czech syntactic rules. However, if the verb is inherently reflexive
(reflexivum tantum), it keeps its single reflexive particle se without adding another.
Example of transformation:
studovat (to study)
ACT(1) PAT(4)
→ (se) ACT(3) PAT(1) adverbial-func(*),
where adverbial-func is typically a modification of manner (MANN).
Example application to a sentence:
Pavel studuje angličtinu. → Angličtina se Pavlovi studuje snadno. (lit. Paul studies
English. → English (itself) to-Paul studies easily.)
3.7
Reciprocity
Strictly speaking, reciprocity is not a diathesis in the proper sense, since the changes on
the surface are not related to the change in voice or other inflectional feature of the verb.
It is considered a diathesis because it shares some of the features of the other types of
diatheses (e.g. adding a particle se to the verb).
In reciprocity, a single entity (or a group of entities, syntactically expressed as a
single phrase, perhaps a coordinated one) occupies two arguments of the verb at the
same time: most often it is the actor (ACT) and patient (PAT), but other pairs are also
possible, e.g. PAT and ADDR.
Except for adding the particle se (unless, similarly to the dispositional modality
type, already inherently present), no other changes in form are visible, except that one
argument (typically, the “later” one of the two arguments being involved) is missing in
the surface and the other one must be either
– in semantic plural (i.e., in surface morphosyntactic plural or it must be a mass noun,
which keeps its surface singular inflection);
– a coordination, typically a conjunction, unless all members of the coordination
fulfill on of the conditions on this list;
– construction of a single noun phrase modified by another prepositional phrase using the preposition s(e) (lit. with); this is often considered just another form of
coordination;
Diatheses in PDT-Vallex
373
– a single noun phrase describing a group of people of objects.
In the sentence representation, the “missing” argument is represented by an artificial
t-lema #Rcp and the appropriate functor (PAT, ADDR, etc.), so that obligatory arguments
are present as required. The double-function argument keeps the first functor (ACT in
case of ACT–PAT reciprocity, PAT in case of PAT–ADDR reciprocity, etc.) and its form
has to fulfill one of the above alternatives. Often, but not necessarily, the adverbials
(sebe/sobě) vzájemně/navzájem (lit. (themselves) mutually), jeden druhého/druhému/...
(lit. (to) each other) or similar are used on the surface to explicitly stress the reciprocity.
Additionally, the forms of location adverbials are slightly changed to reflect the
symmetricality of the event, but the description of such changes in free modifiers is
beyond the scope of this article (and has not been dealt with systematically in the PDT
either).
Example of a transformation:
vidět (to see)
ACT(1) PAT(4)
→ (se) ACT(<sem-plural>1) PAT(-)
where <sem-plural> is the surface form of all possible semantic plurals, and (-) means
there should not be any form corresponding to this argument on the surface (moreover,
the t-lema should be #Rcp).
Example application to a sentence:
Pavel viděl Janu na druhé straně ulice. → Pavel a Jana se viděli přes ulici. (lit. Paul
saw Jane at the other side of the street. → Paul and Jane saw (themselves) across
the street.)
4
Future work
In this arcticle, we have described valency in the PDT-Vallex valency dictionary, concentrating on its formal properties and especially on the change of surface form of the
verb and its arguments when secondary diatheses are used. The report presented here is
not exhaustive; more details can be found in [40].
In the future, we would like to fully formalize the transformation changes not only
for the purposes of checking the consistency of annotation (as it was done in the Prague
Dependency Treebank and its PDT-Vallex dictionary), but also to serve, in a simpler
manner than today, the task of natural language generation, where diathesis and specifically its surface from are of high importance ([31]).
Obviously, we will be extending the PDT-Vallex dictionary further. For example, the
dictionary is now being extended to cover all the verbs in the Czech translation of the
Wall Street Journal portion of the Penn Treebank, which is part of a parallel treebank
project called the Prague Czech-English Dependency Treebank.10 . These extensions
will also necessarily lead to more examples of diathesis, and thus might need extensions
of the surface form transformation rules described in this article.
10
http://ufal.mff.cuni.cz/pcedt
374
Zdeňka Urešová and Petr Pajas
Acknowledgements
We gratefully acknowledge the support of the Czech Ministry of Education through
the grant No. MSM-0021620838, the Grant Agency of Charles University in Prague
through the grant No. GAUK 52408/2008 and the Academy of Sciences of the Czech
Republic through the grant No. 1ET101120503.
References
[1] Apresjan, J. D. (1995). Leksičeskaja semantika. Sinonimičeskie sredstva jazyka.
Moskva.
[2] Babby, L. H. (1998). Voice and Diathesis in Slavic. In Position paper presented
at the Workshop on Comparative Slavic Morphosyntax: State of the Art. Indiana
University, Spencer.
[3] Chrakovskij, V. S. (1981). Diateza i referentnost’. (K voprosu o sootnozenii aktivnych, passivnych, refleksivnych i reciproknych konstrukcij). Zalogovye konstrukcii v raznostrukturnych jazykach, pages 5–38.
[4] Daneš, F. (1971). Větné členy obligatorní, potenciální a fakultativní (Obligatory,
Potential and Optional Constituents of the Sentence). Miscellanea Linguistica.
Acta Universitas Palackiana Olomucensis, pages 131–138.
[5] Daneš, F. and Hlavsa, Z. (1987). Větné vzorce v češtině. Academia, Praha.
[6] Grepl, M. and Karlík, P. (1983). Gramatické prostředky hierarchizace syntaktické
struktury věty . Brno.
[7] Hajič, J. (1998). Building a Syntactically Annotated Corpus: The Prague Dependency Treebank. Karolinum, Charles University Press, Prague.
[8] Hajič, J. and Honetschläger, V. (2003). Annotation Lexicons: Using the Valency
Lexicon for Tectogrammatical Annotation. Prague Bulletin of Mathematical Linguistics (PBML), 79–80:61–86.
[9] Hajič, J., Panevová, J., Hajičová, E., Sgall, P., Pajas, P., Štěpánek, J., Havelka,
J., Mikulová, M., Žabokrtský, Z., and Ševčíková-Razímová, M. (2006). Prague
Dependency Treebank 2.0. Linguistic Data Consortium, Philadelphia, PA, USA.
[10] Hajič, J., Panevová, J., Urešová, Z., Bémová, A., Kolářová, V., and Pajas, P.
(2003). PDT-VALLEX: Creating a Large-coverage Valency Lexicon for Treebank Annotation. In Nivre, J. and Hinrichs, E., editors, Proceedings of The Second Workshop on Treebanks and Linguistic Theories, volume 9 of Mathematical
Modeling in Physics, Engineering and Cognitive Sciences, pages 57–68, Vaxjo,
Sweden. Vaxjo University Press.
[11] Hlaváčková, D., Horák, A., and Kadlec, V. (2006). Exploitation of the VerbaLex
Verb Valency Lexicon in the Syntactic Analysis of Czech. In Lecture Notes in
Artificial Intelligence, Proceedings of Text, Speech and Dialogue 2006, volume
4188, pages 79–86, Berlin and Heidelberg. Springer.
[12] Karlík, P., Nekula, M., and Pleskalová, J. (2002). Encyklopedický slovník češtiny.
Nakladatelství Lidové noviny, Praha.
[13] Karlík, P., Nekula, M., and Rusínová, Z. (2000). Příruční mluvnice češtiny. Nakladatelství Lidové noviny, Praha.
Diatheses in PDT-Vallex
375
[14] Kettnerová, V. and Lopatková, M. (2009). Changes in Valency Structures of Verbs:
Grammar vs. Lexicon. In This Volume, Proceedings of Slovko 2009. SpringerVerlag Berlin Heidelberg.
[15] Levin, B. (1993). English Verb Classes and Alternations: A Preliminary Investigation. University of Chicago Press, Chicago.
[16] Lopatková, M. and Panevová, J. (2005). Valence vybraných sloves pohybu v češtině. In Piper, P., editor, Proceedings of Matica Srpska za slavistiky, pages 1–8,
Novi Sad, Serbia and Montenegro, Oct. 27-29.
[17] Lopatková, M., Žabokrtský, Z., Benešová, V., Skwarska, K., Bejček, E., Chvátalová, K., Nová, M., and Tichý, M. (2007a). Valenční slovník českých sloves.
Karolinum, Praha.
[18] Lopatková, M., Žabokrtský, Z., Benešová, V., Skwarska, K., Bejček, E., Chvátalová, K., Nová, M., and Tichý, M. (2007b). VALLEX 2.5 - Valency Lexicon of
Czech Verbs, version 2.5.
[19] Mel’čuk, I. A. (1987). Dependency Syntax: Theory and Practice. State University
of New York Press, Albany.
[20] Mel’čuk, I. A. (2003). Actants. In Meaning-text theory 2003. Proceedings of the
First International Conference on Meaning-text theory, pages 111–127.
[21] Mikulová, M., Bémová, A., Hajič, J., Hajičová, E., Havelka, J., Kolářová, V.,
Lopatková, M., Pajas, P., Panevová, J., Razímová, M., Sgall, P., štěpánek, J.,
Urešová, Z., Veselá, K., žabokrtský, Z., and Kučová, L. (2005). Anotace na
tektogramatické rovině Pražského závislostního korpusu. Anotátorská příručka.
Technical Report TR-2005-28, Univerzita Karlova v Praze, MFF, ÚFAL, Prague.
[22] Nasr, A. and Rambow, O. (2004). SuperTagging and Full Parsing . In Proceedings of Seventh International Workshop on Tree Adjoining Grammar and Related
Formalisms .
[23] Padučeva, E. V. (2000). Verbs implying semantic role of result: correlation between diathesis and aspectual meaning. Linguistische Arbeitsberichte 75, pages
125–136.
[24] Padučeva, E. V. (2002). Diateza i diatetičeskij sdvig (diathesis and diathesis shift).
Russian Linguistics, 26:179–215.
[25] Pala, K. and Ševeček, P. (1997). Valence českých sloves, pages 41–54. Brno.
[26] Palmer, M., Gildea, D., and Kingsbury, P. (2005). The Proposition Bank: An
Annotated Corpus of Semantic Roles. Computational Linguistics, 31(1):71–106.
[27] Panevová, J. (1975). On verbal frames in functional generative description II.
Prague Bulletin of Mathematical Linguistics, (23):17–52.
[28] Panevová, J. (1980). Formy a funkce ve stavbě české věty. Academia, Praha.
[29] Panevová, J. (2002). Sloveso: centrum věty; valence: centrální pojem syntaxe. In
neuvedeno, editor, Aktuálne otázky slovenskej syntaxe, pages x1–x5.
[30] Pauliny, E. (1943). Štruktúra slovenského slovesa. SAVU, Bratislava.
[31] Ptáček, J. and Žabokrtský, Z. (2006). Synthesis of Czech Sentences from Tectogrammatical Trees. In Lecture Notes in Computer Science, Proceedings of
the 9th International Conference, TSD 2006, number 4188 in Lecture Notes in
Computer Science, pages 221–228, Berlin / Heidelberg. Springer-Verlag Berlin
Heidelberg.
376
Zdeňka Urešová and Petr Pajas
[32] Pustejovsky, J. (1996). Generative Lexicon. MIT Press, Cambridge, Massachussetts.
[33] Sgall, P., Hajičová, E., and Panevová, J. (1986). The Meaning of the Sentence in
Its Semantic and Pragmatic Aspects. Dordrecht, Reidel, and Prague, Academia,
Prague.
[34] Skoumalová, H. (2001). Czech Syntactic Lexicon. Ph.D. Thesis. PhD thesis,
Faculty of Philosophy, Prague.
[35] Štícha, F. (1984). Utváření a hierarchizace struktury větného znaku. Univerzita
Karlova, Praha.
[36] Storrer, A. (1992). Verbvalenz. theoretische und methodische grundlagen ihrer
beschreibung in grammatikographie und lexikographie. Reihe Germanistische
Linguistik, 126:414.
[37] Svozilová, N., Prouzová, H., and Jirsová, A. (1997). Slovesa pro praxi. Valenční
slovník nejčastějších českých sloves. Academia, Praha.
[38] Tesnière, L. (1959). Eléments de syntaxe structurale. Klincksieck, Paris.
[39] Urešová, Z. (2005). Verbal Valency in the Prague Dependency Treebank from the
Annotator’s Point of View, pages 93–112. Veda Bratislava, Slovakia.
[40] Urešová, Z. (in prep.). Valence sloves v Pražském závislostním korpusu. PhD
thesis, Univerzita Karlova v Praze, MFF, Praha.
[41] Uspenskij, V. A. (1977). K ponjatiju diatezy. Problemy lingvističeskoj tipologii i
struktury jazyka, pages 65–84.
[42] Žabokrtský, Z. (2005). Valency lexicon of Czech verbs (PhD thesis). PhD thesis,
Univerzita Karlova v Praze, MFF, ÚFAL, Prague.
A Corpus of Spoken Language and Its Usefulness in the
Research on Language Contact
Marcin Zabawa
Institute of English, University of Silesia, Poland
Abstract. The aim of the present paper will be to discuss two seemingly distant
phenomena, namely (1) spoken language corpora and (2) language contact, or,
to be more precise, the research on the borrowings of English origin in spoken
informal Polish. It would seem that the most objective method of studying
loanwords in spoken language is the use of the corpus, as it enables a linguist to
formulate hypothesis on solid bases. It is then not only possible to state the
existence of a given feature, but also to provide the evidence that would not be
available without a corpus. Unfortunately, however, the corpora of spontaneous
spoken Polish are practically non-existent; as a consequence, the only solution
that seems to be left is to create such a corpus on one’s own. In my paper I
would like to concentrate on discussing the process of collecting and analyzing
a corpus of spoken language which can then be used for the research on English
borrowings. Special emphasis will be placed on various traps and problems
associated with the task in question, such as the choice of material, the choice
of informants, the method of transcribing recorded texts, the (non-)representativeness of the corpus, etc. The present paper is based on the author’s
doctoral dissertation [18].
1 General information about corpus studies
The term ‘corpus’ (form Latin ‘body’) is defined differently by various specialists
working in the field. Some of the definitions are very broad and general, e.g. ‘any
collection of more than one text’ [10], cited also in [11], whereas others are narrow,
but at the same time fairly comprehensive, e.g.
a collection of linguistic data, either written texts or a transcription of recorded speech,
which can be used as a starting-point of linguistic description or as a means of verifying
hypotheses about a language [1].
McEnery and Wilson [10] add that although a corpus can be understood as any
body of text, the term is normally used in modern linguistics in somewhat restricted
sense. In other words, a corpus must have some specific features, described under four
main headings: sampling and representativeness, finite size, machine-readable form
and a standard reference (for details, cf. [10]).
Corpora can be basically divided into three main groups: (1) containing only
written texts, (2) containing only spoken texts (transcriptions) and (3) composed of
both written and spoken language. It seems fairly obvious that it is the spoken
language corpus that is particularly difficult to construct and analyze. According to
Sinclair:
378
Marcin Zabawa
Most corpora keep well away from the problems of spoken language – with some
honourable exceptions – and, for a corpus which in any way purports to reflect a ‘state
of the language’, this is most unfortunate. Many language scholars and teachers believe
that the spoken form of the language is a better guide to the fundamental organization
of the language than the written form; and many writers comment on the differences. In
my own experience, there is no substitute for impromptu speech […] [17].
Nowadays large corpora, stored in computer memory, are an invaluable aid in
many branches of linguistics, most notably in lexicography, as the majority of the new
dictionaries (both mono- and bilingual) are based on corpora, e.g. Collins Cobuild
English Dictionary for Advanced Learners, 3rd edition, ed. by Sinclair (2001), based
on the Bank of English, a corpus consisting of around 400 million words or Oxford
Dictionary of English, 2nd edition, ed. by Soanes and Stevenson (2003), based on the
Oxford English Corpus, composed of the British National Corpus (100 million
words), the Oxford Reading Programme (around 77 million words) and other
databases.
Apart from lexicography, corpora can be successfully used in language teaching,
historical linguistics, pragmatics and discourse analysis, sociolinguistics, stylistics,
and many other spheres (for details, cf. [10]). They can also be used in the studies
connected with language contact.
2 Corpora and language contact
The use of a corpus enables linguists to formulate hypotheses about language on solid
bases. It is thus possible not only to state the existence of a given feature, but also to
provide the evidence that would not be available without a corpus. As for the study of
language contact (manifested through the existence of borrowings), a linguist working
with a corpus typed into a computer is not only able to state the existence of a given
loan in a language, but also to discuss its frequency, both absolute and relative (i.e.
compared to other loans), usually in terms of types and tokens, conduct the contextual
analysis, determine all the senses in which a word was used and rank them according
to their frequency, discuss the relation between the usage of a loan and various other
criteria, such as the types of texts, the topics, the informants (their sex, education, age,
knowledge of foreign languages). Moreover, as Otwinowska-Kasztelanic [12] rightly
notices, it is not only the existence of a given feature that may be linguistically and/or
statistically significant, but the non-existence of it may also constitute an important
observation.
As for the language contact, it can be stated that the influence of English upon
Polish has been researched relatively thoroughly in Polish linguistics. Sadly, however,
most of the studies concentrate on discussing individual instances of borrowings in
Polish (usually based on written sources, e.g. the language of the press) and thus they
do not constitute a systematic research on the subject. The situation is even worse in
the case of the research of English borrowings in spoken Polish, as there are very few
A Corpus of Spoken Language and Its Usefulness in the Research on Language Contact 379
studies on the subject. It would seem that the most objective method of studying
loanwords in spoken language is the use of the corpus, as it enables a linguist to
formulate hypothesis on solid bases. However, the corpora of spoken Polish are
scarce, to say the least. This is the reason why it is in fact necessary to collect such a
corpus on one’s own.
3 Corpora of spoken Polish
As Dunaj notes (cited in [12]), the traditional linguistic studies in Poland were based
solely on written Polish. As a consequence, there was hardly any research on spoken
Polish before the 1960s (or even early 1970s), with the exception of dialectological
studies. A similar view was expressed by Żydek-Bednarczuk [19], who states that the
studies on spoken language were intensified only during the last 25 years, dealing
mainly with phonetic realization, vocabulary, syntax and semantics.
At first, the studies on spoken language were connected with the research on
dialects. It was only in the 1970s that the informal variety of Polish started to be
considered as worthy of serious linguistic investigation. One could mention here such
works as e.g. the study on the syntax of telephone conversation [14], or the research
on text structure of informal conversation [19].
Before extensive recordings of spontaneous spoken Polish were carried out, some
theoretical preliminaries were discussed, such as the typology of spoken Polish
varieties (Buttler, Wilkoń, both cited in [19]), various types of language contact and
language varieties ([6], [7]; cf. also [12]) or components of speech act in a text
(Pisarkowa, cited in [19]). Following those introductory works, studies on urban
Polish were carried out, notably in the following Polish cities: Katowice, Kraków,
Łódź, Poznań and Warszawa (cf. [19], [12]). As a consequence, the following corpora
of spoken Polish, among others, were published:
•
•
•
•
in the region of Katowice: Teksty języka mówionego mieszkańców miast
Górnego Śląska i Zagłębia [7], [8]
in Kraków: Wybór tekstów języka mówionego mieszkańców Krakowa [2]
in the region of Łódź: Wybór tekstów języka mówionego mieszkańców Łodzi
i regionu łódzkiego. Generacja najstarsza [4], Wybór tekstów języka
mówionego mieszkańców Łodzi. Generacja starsza, średnia i najmłodsza [5]
in the region of Warszawa: Korpus języka mówionego młodego pokolenia
Polaków [13].
However, there is still a shortage of corpora of spoken Polish, as the existing ones
are relatively small, e.g. the corpus of informal spoken Polish of the Warsaw variety
[13]. Naturally, there exist larger corpora of English, used primarily in lexicography,
such as Korpus języka polskiego (http://korpus.pwn.pl/) or Korpus IPI PAN
(http://korpus.pl/index.php?page=welcome). The problem is, however, that they cannot be treated as corpora of spoken Polish. In the case of the former
380
Marcin Zabawa
one, only 4.5% of words come from recorded conversations whereas the latter one
does not contain spoken texts at all.
Consequently, as the corpora of spontaneous spoken Polish are practically nonexistent (with the exception of the small corpora listed above), it would seem that –
when one wants to analyze the borrowings used in spoken informal Polish – the only
solution is to collect such a corpus on one’s own. However, certain problems may
(and will) arise. These are discussed in the next section.
4 Problems connected with spoken language corpora
When one decides to construct a corpus of spoken language and analyze it in terms of
borrowed words and/or meanings, three difficulties are predominant: first, it is much
more difficult to construct a representative corpus of spoken language, particularly
consisting of informal conversations, than, say, of the language of the press. On one
hand, to ensure a good quality of the recording and make it in accordance with the
law, it should be carried out openly, i.e. non-surreptitiously. On the other, the
presence of a tape recorder may and often does have influence on the linguistic
behaviour of speakers, so it is necessary to make sure that the speakers do not feel
inhibited and behave as naturally as possible. Moreover, it is often not easy to find
appropriate people that would act as informants. They should be varied as to their sex,
age, education, place of work, social and economic background and so on. Second,
the creation of a corpus of spoken language is a difficult and time-consuming process,
as the conversations must be recorded, transcribed and finally typed into a computer.
The last two phases are particularly time-consuming, especially when compared to the
process of creating a corpus of written language, which does not need to be
transcribed and, instead of keyboarding, one can use faster and more efficient
methods, such as optical scanning, i.e. machine reading (suitable especially for
printed books, as Sinclair [17] notices) or the re-use of the material already in
electronic form (suitable especially for press texts, which can frequently be found on
the Internet). One could of course make use of such quasi-spoken varieties as
interviews or film scripts but, as Sinclair [17] notes, they are not really instances of
spoken, but rather written-to-be-spoken language; hence they do not have typical
features of spontaneous informal conversations and could not be included in the
corpus of spontaneous language:
If it is impossible in an early stage of a project to collect the spoken language, then
there is a temptation to collect film scripts, drama texts, etc., as if they would in some
way make up for this deficiency. They have a very limited value in a general corpus,
because they are ‘considered’ language, written to simulate speech in artificial settings.
Each has its own distinctive features, but none truly reflects natural conversation, which
for many people is the quintessence of the spoken language. There is special integrity in
a text which is a full record of a public meeting, enquiry, court case, radio or television
station, etc., despite the mix of impromptu and considered language that is used –
A Corpus of Spoken Language and Its Usefulness in the Research on Language Contact 381
scripts and even read-out statements are common. But such records are not likely to be
representative of the general usage of conversation. [17]
The above quotation answers the question as to why it was decided to exclude
such texts as interviews or discussions emitted e.g. by the television from the corpus
compiled by the present author.
Third, a typical spoken-language corpus does not contain many borrowed words,
as Poplack et al. pointed out:
Gathering enough data for the systematic study of the use of borrowed words in a
speech community is inherently very difficult. Tokens of these words are typically rare
in monolingual discourse, so that several hours of speech will yield only a few dozen,
most of which occur only once. In certain contexts, of course, and for certain topics of
conversation, there will be some set of borrowings which are used repeatedly, but the
imposition of contextual or topical restrictions would vitiate the comprehensiveness and
representativity of any investigation attempting to give a general characterization of
borrowing and integration. [16]
A similar view was expressed by Dunaj:
Każdy, kto zajmował się badaniami języka mówionego, wie, że uzyskanie obfitych
materiałów leksykalnych z wypowiedzi mówionych nie jest łatwe. Wymaga żmudnych,
długotrwałych obserwacji. [3]
Any person conducting research on spoken language knows that it is not easy to obtain
rich lexical material from spoken utterances. It requires long and laborious
observations. [my translation]
The above quotations answer the question as to why corpora of spoken language
are infrequently used for the systematic study of loans in a language. Instead, as
Poplack et al. [16] notes, most studies dealing with the use of borrowed words in
spoken language resort to three alternatives:
(1) Artificial methods of the elicitation of data, used by e.g. Poplack and Sankoff
[15], who prepared a random series of photographs of 45 everyday items which could
be designated by concrete nouns. The bilingual (Spanish-English) informants were
then asked to name the object and provide any additional words for the same concept
they could think of.
(2) The analysis of a few isolated borrowings that occurred naturally, a method
used by e.g. Mougeon et al., cited in [16].
(3) The analysis of anecdotal lists of borrowed words, a method used by e.g. Nash,
cited in [16].
Nevertheless, it seems that such methods as e.g. using photographs to elicit nouns
from the respondents can provide only indirect data and thus the results obtained in
this way are not necessarily representative of the use of loanwords in informal
spontaneous conversations. As a consequence, it was decided for the purpose of my
doctoral dissertation to gather a corpus of spontaneous conversations, since there is
still a shortage of studies dealing with the use of borrowings, particularly semantic
ones, in spoken discourse. Some of the recordings included in the corpus (6 out of 20)
382
Marcin Zabawa
are connected with the topic of computers, the Internet or modern technology so as to
ensure that a larger number of lexical and semantic loans will be available for analysis
(cf. also Section 5.3 dealing with the topics of the recordings).
5 The corpus compiled by the author
5.1 Introduction
To ensure the reliability and homogeneity of the corpus, it was decided that only
natural spontaneous conversations will be included. This automatically excludes not
only all written-to-be-spoken forms (such as film dialogues) but also conversations
conducted in a formal setting, such as interviews, discussions or talk shows emitted
by the television.
In general, the corpus is similar to the use collected by Otwinowska-Kasztelanic
[13]. Her corpus, however, consists of conversations conducted in Warsaw whereas
the one compiled by the present author consists of the ones recorded in Upper Silesia.
The entire corpus consists of twenty recordings (60,564 running words altogether).
One could possibly state that the size of the corpus is too small to draw general
conclusions concerning the use of lexical and semantic loans in spoken Polish. It
seems, however, that it is large enough to highlight certain tendencies. Besides, the
entire corpus was collected, transcribed and analyzed by only one person, namely the
author of the present paper. As a consequence, it could not be very large for practical
reasons.
5.2 The informants
Altogether, 48 people (including the author of the study) participated in the
conversations (31 women and 17 men). They are uniquely coded throughout all the
recordings. In other words, the same person is given the same symbol in all the
conversations in which he or she took part. The codes for female speakers are F1, F2,
…, F30, F31 whereas the codes for male ones are M1, M2, …, M16, M17.
It must be noted at this point that nine speakers were not taken into account in the
present study. They were accidental speakers (e.g. people asking for something), who
uttered only a few words altogether. The number of speakers taken into consideration
is thus 39 (25 women and 14 men). The basic information about the speakers
(gathered in 2003) is presented below. The following information is given: the symbol
of the respondent, his/her age, sex, level of education (and the field of education in
brackets) and his/her knowledge of English. The knowledge of English of most of the
informants was established by means of informal conversations and/or short
placement tests. With this end in view, a four-point scale was used: none, basic,
intermediate and advanced. It must be added that most of the speakers from the first
group know a few English words and expressions, at least passively. The last group,
in turn, includes also informants fluent in English.
A Corpus of Spoken Language and Its Usefulness in the Research on Language Contact 383
M1
M2
M3
M4
M5
M6
M7
M8
M9
M10
M11
M12
M13
M14
M15
M16
M17
F1
F2
F3
F4
F5
F6
F7
F8
F9
F10
F11
F12
F13
F14
F15
F16
F17
F18
F19
F20
F21
F22
F23
F24
F25
F26
F27
F28
F29
F30
F31
26
male
higher (physical education)
36
male
higher (history)
25
male
higher (English studies)
25
male
higher (physical education)
not taken into account in the present study
26
male
higher (English studies)
23
male
higher (computer science)
54
male
higher (law)
38
male
secondary (technical education)
not taken into account in the present study
49
male
secondary (technical education)
not taken into account in the present study
24
male
secondary (technical education)
27
male
higher (computer science)
26
male
higher (physical education)
50
male
secondary (technical education)
22
male
secondary (general education)
26
fem.
higher (psychology)
40
fem.
higher (Russian studies, English studies)
25
fem.
higher (biology)
27
fem.
higher (English studies)
55
fem.
higher (biology)
46
fem.
higher (history)
29
fem.
higher (psychology, English studies)
not taken into account in the present study
25
fem.
higher (biology)
29
fem.
higher (Polish studies)
31
fem.
higher (theology)
29
fem.
higher (Polish studies)
35
fem.
higher (chemistry)
not taken into account in the present study
36
fem.
higher (library science)
27
fem.
higher (German studies)
48
fem.
higher (German studies)
not taken into account in the present study
not taken into account in the present study
not taken into account in the present study
69
fem.
primary (general education)
36
fem.
vocational (gastronomy)
not taken into account in the present study
47
fem.
vocational (technical education)
51
fem.
higher (medicine)
23
fem.
vocational (hairdressing)
23
fem.
vocational (gastronomy)
24
fem.
vocational (hairdressing)
35
fem.
higher (geography)
50
fem.
higher (law, English studies)
24
fem.
higher (marketing and management)
intermediate
basic
advanced
basic
advanced
intermediate
none
none
basic
none
basic
intermediate
none
basic
intermediate
advanced
basic
advanced
none
none
advanced
intermediate
intermediate
basic
intermediate
none
basic
basic
basic
none
none
none
basic
basic
basic
none
basic
advanced
intermediate
384
Marcin Zabawa
As one can see, most of the informants are university graduates, but there is also a
small percentage of people having secondary or vocational education. The age of
most of the speakers (27 out of 39) ranges between 21 and 36.
To make the corpus conversations as varied as possible, it was decided to choose
respondents representing various occupations, including the following trades and
professions: primary and secondary school teachers (of various subjects), a university
teacher, a computer programmer, computer specialists, office workers, an entrepreneur, an interpreter, a lawyer, a bank clerk, security guards, an electrician, a lorry
driver, an electronics engineer, shop assistants, a psychologist, a cleaner, a cook, a
doctor (physician), hairdressers, an old-age pensioner and unemployed persons. All of
the speakers come from Upper Silesia in Poland.
5.3 The recordings
Altogether, there were twenty conversations recorded for the purpose of the study.
Most of the conversations were recorded in informal situations, happening both
indoors, during such events as a birthday party, a family meeting or a meeting with
friends, cf. Recordings 5, 12 and 20 (12,022 words in total) and outdoors, during
walks, often with a dog, e.g. in a park or the woods, along the lake shore, through a
housing estate etc., or during informal meetings with friends and/or family members
e.g. in a garden, cf. Recordings 1, 6, 9, 14, 16, 18 and 19 (25,871 words in total).
Some other recordings, viz. number 2, 10 and 13 (6,581 words in total), were
conducted in a place of work, namely in a staffroom in a school during long breaks.
The situation was thus more formal than in the previous cases, but nevertheless the
conversations could be safely described as informal, as most of the teachers taking
part in them were on first name terms with one another. The rest of the recordings,
namely number 3, 4, 7, 8, 11, 15 and 17 (16,090 words in total) were conducted
between the students learning German in a foreign language school. The
conversations were spontaneous and informal (recorded before the actual classes), as
all the people involved were again on first name terms.
The topics of particular recordings and the number of running words and are
presented below. The sequence of information is as follows: the number of the
recording, the topic or topics of the conversation, the number of running words in a
given recording.
1
2
3
4
5
6
7
8
9
dogs, holidays, sport contests, fishing
school, teachers, pupils
computers, the Internet, computer programs
student exchange, computers and e-mail, exams and cheating
acquaintances, family matters, working, excursions, holidays
computers, computer games, computer magazines, the Internet
computers and e-mail, student exchange, learning German
teaching and learning foreign languages
a wedding, doing shopping, business matters
3966
2868
1620
2027
3290
3654
2348
3144
3327
A Corpus of Spoken Language and Its Usefulness in the Research on Language Contact 385
10
11
12
13
14
15
16
17
18
19
20
school, pupils and teaching, doctors and health-care
correspondence, computers (esp. using a text processor)
family matters, acquaintances, keeping dogs at home
films, teaching and giving grades, marriages
looking for work, doing business, computers and the Internet
weather, holidays
looking for work, learning abroad, moving to another country
learning foreign languages, taking exams
birthdays, films, new technologies, software piracy
grilling and smoking food, working, ticks and other insects
everyday life, cooking and eating, family, taking care of a baby
2178
2393
3848
1535
4060
2229
3895
2329
3807
3162
4884
As one can see, it was decided to include conversations on various topics. Three of
them, however, seem to be prevailing: (1) everyday activities, such as working, living,
talking about one’s family, friends and/or acquaintances (Recordings number 5, 9, 12,
19, 20), (2) computers and the Internet (Recordings number 3, 4, 6, 7, 11, 18) and (3)
teaching and learning (Recordings number 2, 8, 10, 13, 17). The conversations about
computers have turned out to be particularly interesting for the study, as they contain
a large percentage of both lexical and semantic loans found in the corpus (compared
to other recordings).
5.4 The type of language used
As was noted above, the entire corpus has been composed of spontaneous
conversations, hence the language used in it can be safely described as naturally
occurring Polish. It should also be underlined that the language of some of the
speakers has visible features of the Silesian dialect (notably the one of F21, F22, F24,
F25, F28, M6, M7, M8, M9, M11, M13 and M16). It must be added that the situation
of most of the speakers taking part in the study can be characterized as diglossic; in
other words, the informants are able to speak both standard Polish and the Silesian
dialect and alternate between them according to the situation. Some recordings are
thus examples of naturally occurring Silesian dialect (especially number 1, 5, 6, 9, 12,
14, 18 and 19).
For the sake of clarity, only some most noticeable features of the Silesian dialect
are marked in the corpus, e.g. some aspects of the special pronunciation (such as
special endings or extra consonants), some untypical aspects of morphology or the use
of specific vocabulary.
5.5 The transcription and the use of fonts
As the aim of the research was to study lexical and semantic loans in informal spoken
Polish, it was not necessary to transcribe the recorded conversations phonetically.
Instead, it was decided to use orthographic transcription. However, no punctuation
marks were used, with the exception of a question mark, used for marking questions
(indicated by the intonation used by a speaker). Moreover, the slash symbol ( / ) was
386
Marcin Zabawa
used to show short (up to one second) pauses in the speech. Such a notation is aimed
at reflecting the natural flow of speech. The convention used for the transcription of
the recordings used for the purpose of the present study is thus very similar to the one
used by Żydek-Bednarczuk [19], who conducted a study on the structure of informal
conversation.
A speaker’s turn is always indicated by his or her code, i.e. a letter (F for women,
M for men) followed by a number and a colon, all written in bold for easy reference.
All names, surnames, nicknames and some other expressions denoting people (with
the exception of some of them referring to celebrities) are replaced by common nouns
given in curly brackets (e.g. {name}, {surname}, {nickname}, {company name},
{city}) in order to keep the informants and the people they speak about anonymous.
Other features marked in the transcriptions include: the pronunciation of items
borrowed from English, given in angle brackets; non-verbal behaviour of the
speakers, such as laughter, a cough, a whistle, etc., with the indication of its duration
(in general, the duration of the phenomena lasting up to one second is marked as
‘[1 sec.]’) given in square brackets; semi-verbal behaviour, expressing some kind of
indecision, tacit agreement or having no real semantic function, marked as mmm, yyy,
mhm; whisper; background noises (e.g. the sounds of various machines or appliances),
together with their duration, indicated in square brackets; pauses in speech, with the
indication of their duration; overlapping speech and moments of unclear speech
(indistinguishable or incomplete words). Besides, the English lexical and semantic
borrowings were marked (by means of the use of bold type for lexical borrowings and
underlining for semantic borrowings).
6 The results of the study – a brief summary
The results of the study, i.e. the use of English borrowings in informal spoken Polish,
will not be discussed in detail, since the aim of present article was to describe the
corpus collected for the study as well as to present certain problems associated with
collecting a corpus of spontaneous spoken language.
Altogether, 78 English lexical borrowings (including derivatives) were found in the
corpus (225 tokens). This number may seem large at first glance, but when it is contrasted
with the number of running words of the entire corpus, it becomes evident that English
loanwords constitute a very small percentage of the corpus, namely 0.3715%. As for
semantic borrowings, 44 types of them (including derivatives) were found in the corpus
(158 tokens), which constitutes 0.26088% of the entire corpus [18].
7 Conclusions
Altogether, it turned out, as was expected, that the use of a corpus is not the most
efficient way of studying English borrowings in informal speech, since the loans (both
lexical and semantic) are not very numerous. What is more, the majority of them were
A Corpus of Spoken Language and Its Usefulness in the Research on Language Contact 387
used only once in the entire corpus. On the other hand, many of them have not been
discussed in the literature on the subject so far. Most importantly, however, it would
seem that the study can be described as a reliable one: the conclusion is that English
borrowings are used relatively infrequently in spoken informal Polish.
In conclusion, it can also be stated that there is still a need to study English
borrowings in spoken Polish, as this area is generally neglected. However, the project
of this type should not be carried out by an individual linguist, but by a number of
them, preferably from different universities (i.e. located in different parts of Poland)
so that the informants would be varied as to their place of living, sex, age, occupation,
education, social and economic background and knowledge of English and other
foreign languages. The base of such a research should be a large corpus of spoken
spontaneous language (i.e. forms such as discussions, interviews emitted e.g. by the
television should not be included), as it remains the only objective method of studying
loanwords in a spoken variety of a language.
References
[1] Crystal, D. (1997). A Dictionary of Linguistics and Phonetics (4th edition).
Oxford: Blackwell Publishers Ltd.
[2] Dunaj, B. (ed.) (1979). Wybór tekstów języka mówionego mieszkańców
Krakowa. Kraków.
[3] Dunaj, B. (2000). Nowe słownictwo w leksykografii. In: Mazur, J. (ed.)
Słownictwo współczesnej polszczyzny w okresie przemian, pages 33–38,
Lublin: Wydawnictwo Uniwersytetu Marii Curie-Skłodowskiej.
[4] Kamińska, M. (ed.) (1989). Wybór tekstów języka mówionego mieszkańców
Łodzi i regionu łódzkiego. Generacja najstarsza. Łódź.
[5] Kamińska, M. (ed.) (1992). Wybór tekstów języka mówionego mieszkańców
Łodzi i regionu łódzkiego. Generacja starsza, średnia i najmłodsza. Łódź.
[6] Lubaś, W. (ed.) (1976). Miejska polszczyzna mówiona. Metodologia badań.
Katowice.
[7] Lubaś, W. (ed.) (1978). Teksty języka mówionego mieszkańców miast Górnego
Śląska i Zagłębia, vol. 1. Katowice.
[8] Lubaś, W. (ed.) (1980). Teksty języka mówionego mieszkańców miast Górnego
Śląska i Zagłębia, vol. 2. Katowice.
[9] Lubaś, W. (1979). Społeczne uwarunkowania współczesnej polszczyzny. Szkice
socjolingwistyczne. Kraków: Wydawnictwo Literackie.
[10] McEnery, T. and Wilson, A. (1996). Corpus linguistics. Edinburgh: Edinburgh
University Press.
[11] Myrczek, E. (2000). Corpus – its definition, compilation, taxonomy and future.
Linguistica Silesiana 21, 43–62.
388
Marcin Zabawa
[12] Otwinowska-Kasztelanic, A. (2000). A study of the lexico-semantic and
grammatical influence of English on the Polish of the younger generation of
Poles (19-35 years of age). Warszawa: Wydawnictwo Akademickie Dialog.
[13] Otwinowska-Kasztelanic, A. (2000). Korpus języka mówionego młodego
pokolenia Polaków. Warszawa: Wydawnictwo Akademickie Dialog.
[14] Pisarkowa, K. (1975). Składnia rozmowy telefonicznej. Wrocław: Zakład
Narodowy im. Ossolińskich.
[15] Poplack, S. and Sankoff, D. (1984). Borrowing: the synchrony of integration.
Linguistics 22, 99-135.
[16] Poplack, S., Sankoff, D. and Miller, Ch. (1988). The social correlates and
linguistic processes of lexical borrowing and assimilation. Linguistics 26, 47104.
[17] Sinclair, J. (1991). Corpus, Concordance, Collocation. Oxford: Oxford
University Press.
[18] Zabawa, M. (2006). English lexical and semantic loans in informal spoken
Polish. Vol. I-II. (Vol I: The dissertation; vol. II: The corpus of informal
spoken Polish). Unpublished doctoral dissertation. University of Silesia.
[19] Żydek-Bednarczuk, U. (1994). Struktura tekstu rozmowy potocznej. Katowice:
Wydawnictwo Uniwersytetu Śląskiego.
Vybudování databází na základě slovníku jako korpus
Miloud Taïfi1 a Patrice Pognan2
1
Universita Sidi Mohamed ben Abdallah, Fez, Maroko
2
INALCO, LALIC-Certal, Paříž, Francie
Abstrakt. Slovník berberštiny středního Maroka (tamazight) obsahuje skoro
7200 kořenů. Automatické vybudování databáze se ukazálo možným díky
tomu, že je tento slovník dobře strukturován. Dva moduly jsou spojené
s automatickým vyhledáváním a rozšiřováním struktur, třetí s vybudováním
databází čtením výsledků. Hlavní aplikace bude příprava francouzskoberberského slovníku. Francouzský překlad tohoto článku dáme na server
http://www.lalic.paris4.sorbonne.fr.
První slovník berberštiny středního Maroka (tamazight) byl publikován profesorem
Taïfim v roce 1991 jako rukopis. Druhá verze vyjde koncem roku 2009. Byla
vyhotovena v rámci marocko-francouzské vědecké kooperace. Jedná se o opravenou
a o 50% rozšířenou verzi. Jak to bývá v semitských jazycích, tento slovník je
abecedně seřazen podle kořenů (je jich skoro 7200). Slovník se představuje jako
souhrn 29 souborů ve Wordu (optické snímání rukopisu selhalo a bývalý slovník se
musel ručně naklepat). Je dobře strukturován, jak je z ukázky vidět:
KOŘEN – ♦ slovo – gramatické údaje – ► významy – ● příklady
První etapa spočívala v tom, že se soubory ve Wordu změnily na soubory
v Unikódu UTF-8 a spojily dohromady do poměrně velkého korpusu o 2 700 000
znaků. Na základě tohoto korpusu se vytvořily dva seznamy: korespondence
berberských slov s kořeny a seznam francouzských významů s odpovídajícími
berberskými kořeny.
390
Miloud Taïfi a Patrice Pognan
Třetí a aktuální úkol, stále na základě slovníku, je automatické vybudování
databáze berberštiny středního Maroka. Dva moduly jsou spojené s automatickým
vyhledáváním struktur, třetí s vybudováním databází:
1 automatické vyhledávání slovníkových struktur
1.1 rozpoznávání aktuálních struktur slovníku
1.2 rozšiřování struktur doplňováním morfologických údajů
2 vybudování databází čtením výsledků z etapy 1.2
1 Automatické vyhledávání slovníkových struktur
1.1 Etapa rozpoznávání struktur slovníku je sama rozdělena do pěti programů.
Vychází z takového souboru:
1. První program čte základní soubor v Unikódu, určuje kořeny podle písma
(velká písmena s diakritickými znaménky) a rozděluje jednotlivá hesla prázdným
řádkem. Dále začíná rozeznávat strukturu a označuje ji určitým počtem tabelací
(jedna tabelace pro každou úroveň). Tady určuje úroveň slov.
2. Druhý program jde hlouběji do struktury a určuje významy.
3. Třetí určuje příklady.
4. Tento program provádí změny, které budou během druhé etapy nutné pro
automatické generování morfologických údajů: dává do jednoho záznamu slovo
a základní morfologické údaje.
5. Poslední program lehce upravuje úroveň kořene: přidává k němu odkazy na
případnou přítomnost kořene v jiných jazycích.
Vybudování databází na základě slovníku jako korpus
391
Výsledky první etapy (1.1) jsou uvedeny níže:
1.2 Rozšiřování struktur doplňováním morfologických údajů
Druhá etapa má za úkol automaticky vypočítat další morfologické hodnoty na
základě původní struktury. Je rozdělena do tří programů seřazených podle jevů
a jejich složitosti.
1. V berberštině existuje velká homografie kořenů. Pro přehlednost a pro
zajištění správného fungování databází bylo nutno očíslovat kořeny. Tak např.
kořen „BDR“ uvedený výše bude označen jako „BDR
BDR3“.
2. Vyhledávání lexikálních kategorií a obohacování morfologických údajů
(gramatémů): proveditelnost projektu čistě závisí na tomto programu. Kdyby býval
neprovedl správnou analýzu lexikálních kategorií, byla by konstrukce databází
nemožná.
Neurčuje jenom „třídy“ (lexikální kategorie), ale určuje taky „kmeny“ (typ
odvození: verbo-nominální nebo nominální derivace). Pojem rozdělení lexika na
dvě části, verbo-nominální a nominální, je v lingvistice hamito-semitských jazyků
běžný. Aplikujeme ho i na indoevropské jazyky.
Program rozeznává:
- jednoduchá slovesa označená jako verbo-nominální derivace (VN), sloveso (V)
a jednoduché sloveso (1.1). Číselná škála (1.1. - 1.2. - 1.3. - 2., přitom „1“ odkazuje
na verbo-nominální derivaci a „2“ na nominální derivaci) umožňuje automatické
seřazení slov pod kořeny u výstupů z druhého typu databází (viz níže 2.2.1).
392
Miloud Taïfi a Patrice Pognan
- odvozená slovesa: „VN
V
1.2“
- deverbativa: „VN SUBST 1.3.“
- a konečně „nomen“ (podstatná a přídavná jména) s nominální derivací (N)
s hodnotami (SUBST nebo ADJ) a (2.): „N
SUBST 2“.
U všech sloves (jednoduchých a odvozených) jsou ještě uvedeny čtyři hlavní
časy, které pak umožní automatické časování uvnitř databáze. To jsou (aorist,
dokonavý tvar [také často nazývaný „préteritum“], nedokonavý tvar, negativní
dokonavý tvar) např.:
sloveso „dělat“
:
„sker sker
teskar ur-skir“.
U všech nominálních forem (deverbativa a substantiva) se generuje rod podle
formy slova: až na výjimky, často odvozené z arabštiny, je berberský „nomen“
pravidelný. V kategorii substantiv slova začinající na „a/i/u“ jsou mužského rodu,
slova začinající na „t“ jsou ženského rodu (a většinou končí též na „t“). Střední rod
neexistuje. Dále se generují tzv. „vázané“ formy v jednotném a v množném čísle na
základě „volných“ forem. U přídavných jmen se tyto formy generují pro mužský
a ženský rod.
3. Příklady představují nejhlubší úroveň struktury slovníku. Vedle příkladu
obsahuje ještě jejich záznam překlad a eventuelně doslovný překlad v závorkách.
Struktura záznamu vypádá takto:
∙ irḍel a-s leflus (litt. il a prêté à lui de l’argent), il lui a prêté de l’argent.
„∙ berberský příklad (překlad doslova) , překlad .“
Typický oddělovač je čárka, která odděluje příklad a doslovný překlad v závorkách na jedné straně a korektní překlad na straně druhé. Problém je v tom, že čárka
je špatným oddělovačem proto, že se může navíc objevovat ve všech třech částech
záznamu, a to několikrát. Z toho vyplývá, že odpovídající algoritmus je poměrně
složitý. Nedá se vyloučit, že některé záznamy nebudou plně nebo vůbec vyřešeny.
Jeden z použitých postupů je rozdělení na segmenty ohraničené čárkou a otestování
přítomnosti v segmentu buď typických berberských znaků nebo typických francouzských znaků, a to tak, že berberské znaky se otestují zepředu od druhého segmentu
a francouzské naopak zezadu od předposledního segmentu. Bohužel se může stát,
že jeden segment (ba i několik) uprostřed záznamu nemá vůbec žadný zvláštní
znak!
Zvláštní berberské znaky jsou: „ č š ž ḍ ġ ḥ ṛ ṣ ṭ ẓ ε “ a francouzské jsou: „ ’ à æ
é è ê ï î ô œ ù û ç “.
Uvažujeme také o tom, že budeme vyhledávat v segmentech případnou
přítomnost berberských předložek, spojek a částic jako např. „d“, „n“, „s“ (má
stejný význam jako v češtině a slovenštině!), …
Vybudování databází na základě slovníku jako korpus
393
Výsledky druhé etapy (1.2) vypadají takto:
2 Vybudování databází čtením výsledků z etapy 1.2
Připravili jsme a otestovali dva prototypy. První byl připraven tak, aby odpovídal
struktuře:
KOŘEN – ♦ slovo – gramatické údaje – ► významy – ● příklady.
Druhý má takovou strukturu, kde nejvyšší úroveň je úrovní slova se zmínkou
kořene:
slovo a kořen – lexikální kategorie a gramatické údaje – významy – příklady.
Načteme připravené údaje z etapy 1.2 do těchto dvou prototypů, ale pravděpodobně
budeme nadále pracovat s druhým prototypem, protože odpovídá naší nynější
koncepci databází.
394
Miloud Taïfi a Patrice Pognan
2.1 První prototyp
Jde v podstatě o přesný obraz slovníku:
2.2 Druhý prototyp
Po vybudování několika databází jsme nabyli přesvědčení, že konstrukci databází
do nekonečna opakovat nelze. Proto jsme se rozhodli definovat prototyp, který by
odpovídal našim požadavkům. Vybudovali jsme prototyp databáze, který je
jedinečný, víceúčelový a vícejazyčný.
Tento prototyp odpovídá nejširšímu popisu lingvistického systému indoevropských a hamito-semitských jazyků. Jednoduchou kopií jsme z něho už dostali
několik databází: jednak pro češtinu, slovenštinu, slovinštinu a albánštinu, jednak
pro berberštinu. Rýsuje se i projekt rozpracování všech druhů maghrebské
arabštiny. Prototyp nyní rozšiřujeme na struktury tureckých jazyků hlavně na
základě turečtiny a uzbečtiny. Je to důležité proto, že se tím otevírají další možnosti
směrem ke všem aglutinačním jazykům, jako jsou třeba ugrofinské jazyky,
baskičtina a korejština...
Víceúčelovost prototypu je dána tím, že umožňuje shromáždit lexikální znalosti,
které vyústí ve výrobu slovníků, gramatické znalosti ze všech úrovní lingvistického
popisu, ať jsou synchronické nebo diachronické, a odborné znalosti (v technických,
vědeckých a lékařských oborech). Tato víceúčelovost se odráží v přítomnosti
několika druhů struktur.
Vybudování databází na základě slovníku jako korpus
395
2.2.1 Lexikální struktura je hlavní a centrální strukturou databázového systému.
Je určena především pro výrobu všelijakých rejstříků a slovníků. Podrobný popis se
najde v [15] a [16]. Obsahuje tyto úrovně:
1. slovo a kořen (tento typ databáze má třídění podle slov). Hodnoty (1.1, ... až
2.) vygenerované během etapy 1.2 mají za úkol správné třídění slov u každého
kořene, pokud chceme dostat slovník tříděný podle kořenů.
2. lexikální kategorie (může jich být víc)
3. významy
4. příklady a další...
2.2.2 Přilehlá struktura morfologických údajů
Obsahuje informace o cizím původu, o neologii, o morfématickém rozboru s udáním
kořene a dalších afixů. Podává též všeobecné údaje o slovesech, podstatných
a přídavných jménech.
2.2.3 Vedlejší struktury
Jsou trojího druhu. Jedná se pokaždé o nový směr aplikací.
- tabulky pro časování sloves a skloňování nominálních forem. Mají velkou platnost
pro výuku berberštiny (nejenom v cizině pro cizince a pro děti emigrantů, ale i
v maghrebských školách a sdruženích).
- tabulky pro srovnávání jazyků stejné skupiny. Typickým příkladem je srovnávání
- západních slovanských jazyků, slovinštiny a ruštiny na základě diachronické
fonologie. V berberštině se srovnávací práce týkají hlavně, ba výlučně
synchronického stavu jazyka (výjimku tvoří jedině Allati [2]). Zato existuje až příliš
velká bohatost dialektů a místních variant.
- tabulky pro popis klasifikačních oborů:
Tato složka se týká vědeckých, technických a lékařských oborů, kde klasifikace
hraje velkou roli. V prototypu jsou předem definovány dvě stálé tabulky: latinská
396
Miloud Taïfi a Patrice Pognan
a francouzská. Při budování prototypu jsme tuto složku definovali na základě říše
„funghi“, totiž hub, které již nepatří do rostlinné říše a které se staly samostatnou
říší. Slova, která patří do těchto oborů, budou mít vedle normálního lingvistického
popisu ještě odborný popis v rámci příslušné tabulky. Bude to významná pomoc při
tvorbě terminologických slovníků.
2.3 Automatické vytvoření a rozšíření databáze
2.3.1 Automatické vytvoření databáze
Pokud chceme automaticky vytvořit databázi podle prvního prototypu, stačí si
programem načítat výsledný soubor (etapa 1.2) podle vzoru potřebných tabulek
v předurčeném pořadí:
KOŘEN – ♦ slovo – gramatické údaje – ► významy – ∙ příklady.
Oproti tomu je nutno pro výstavbu databáze podle druhého prototypu smísit dva
první záznamy do jednoho záznamu, a to následujícím způsobem:
Dále je způsob čtení dalších tabulek u obou modelů stejný.
2.3.2 Interní programování databáze
Stejně jako v projektu pro slovanské jazyky je jedním z hlavních bodů automatické
časování a skloňování vedené v rámci databáze. Výsledky jsou pak vloženy do
příslušných tabulek.
Vybudování databází na základě slovníku jako korpus
397
Na příkladu posledního formuláře si ukažeme postup naplňování databáze
programováním. Např. pro záznamy „badr“ a „tubadr“ dává program čtení
základní hodnoty, respektive „badr – badr – tbadar / tbidir – ur-badr“ a „tubadr –
tubadr – ttubadar – ur-tubadr“, do připravených polí s názvem FV1 [forme verbale
1] až FV4 [forme verbale 4] (klasičtější pojmenování jsou vedle polí nebo nad
nimi). Teprve na základě těchto informací se automaticky časuje celé sloveso (to se
dělá v programu Access pomocí objektově orientovaného programovacího jazyka
VBA) a výsledky jsou automaticky vloženy do příslušné vedlejší tabulky (viz
předešlý formulář).
2.3.3 Obsahové rozšíření
Databáze, která se těmito procesy vytvoří, bude dále rozšířena dalšími vědeckými
pracemi, hlavně universitními doktoráty různých autorů, ke kterým patří mimo jiné
Ameur, Amrani, El Mandour, Jarmouni,...
Využití databáze
Tato databáze bude mít různé aplikace v oboru lexikálních a gramatických studií.
Nejdůležitější aplikací bude příprava tvorby francouzsko-berberského slovníku.
Vyvinuli jsme proceduru, která dovede obrátit celý obsah databáze tak, že autor
slovníku z francouzštiny do cizího jazyka má k dispozici veškerý materiál
předešlého slovníku (z cizího jazyka do francouzštiny). Tento postup nemá žádný
vliv na koncepci slovníku, ale zato tvorbu slovníku značně urychluje. Touto
procedurou jsme poměrně rychle vytvořili francouzsko-slovenský slovníček v rámci
evropského projektu na vypracování vyučovací metody slovenštiny [5].
Najdou se i další aplikace jako je gramatika berberštiny a metoda výuky
berberštiny ve francouzštině. Věříme, že taková databáze bude pro výuku
berberštiny velmi užitečná.
Pokud nám to vyjde, dáme konečný výsledek na server pro veřejné použití.
Literatura
[1] Achab, R. (1996). La néologie lexicale berbère (1945-1995). Peeters,
Louvain.
[2] Allati, A. (2002). Diachronie tamazighte. Université de Tétouan.
[3] Ameur, M., Bouhjar, A., Boukhris, F., Boukous, A., Boumalk, A.,
Elmedlaoui, M., Iazzi, E. M., Souifi, H. (2004). Initiation à la langue
amazighe. Publications de l’IRCAM, Rabat.
[4] Ameur, M., Boumalk, A. (2004). Standardisation de l’amazighe. Publications
de l’IRCAM, Rabat.
[5] Baranová, E., Křečková, V., Lemay, D., Pognan, P. (2007). Découvrir et
pratiquer le slovaque. Projet européen ALPCU [ lingua II ], L’Asiathèque,
Paris.
[6] Bentolila, F. (1981). Grammaire fonctionnelle d’un parler berbère. Aït
Seghrouchen d’Oum Jeniba, Selaf, Paris.
398
Miloud Taïfi a Patrice Pognan
[7] Boumalk, A. (2003). Manuel de conjugaison du tachelhit, L’Harmattan,
Paris.
[8] Chaker, S. (1995). Linguistique berbère. Etudes de syntaxe et de diachronie,
Peeters, Louvain.
[9] Dallet, J.-M. (1982). Dictionnaire Kabyle – Français, Selaf, Paris.
[10] El Mountassir, A. (2003). Dictionnaire des verbes tachelhit – français.
(parler berbère du sud du Maroc), L’Harmattan, Paris.
[11] Mammeri, M. (1992). Précis de grammaire berbère. Editions EnAP et Inna
yas, Alger et Tizi-Ouzou.
[12] Naït-Zerrad, K. (1995). tajeṛṛumt n tmaziγt tamirant (taqbaylit). ENAG,
Alger.
[13] Naït-Zerrad, K. (1998 - ). Dictionnaire des racines berbères (formes
attestées). Peeters, Louvain.
[14] Naït-Zerrad, K (2004). Linguistique berbère et applications, L’Harmattan,
Paris.
[15] Pognan P. (2009). Définition d’une base de données linguistiques "à tout
faire" comprenant le français – comme langue pivot – et les langues slaves
de l’Ouest. In Congrès international Studia Romanistica Beliana, Banská
Bystrica.
[16] Pognan P. (2009). Définition d’un prototype général de bases de données
(étude des langues slaves de l’Ouest dans une visée multilingue). In Metalanguage and Encoding Scheme Design for Digital Lexicography –
Innovative Solutions for Lexical Entry Design in Slavic Lexicography,
MONDILEX Third Open Workshop, Académie des Sciences Slovaque,
Bratislava.
[17] Sadiqi, F. (2004). "A Grammar of Amazigh", Université Sidi Mohamed ben
Abdellah-PARS, Fez.
[18] Taïfi, M. (1991). Dictionnaire tamazight – français. Parlers du Maroc
central, L’Harmattan-Awal, Paris.
[19] Taïfi, M. (2006). Si les berbérophones ne s’entendent pas oralement, qu’ils
s’écrivent ! Pour une écriture grammaticale du berbère. In Standardisation
de l’amazighe. Publications de l’IRCAM, Rabat.
[20] Taïfi, M., Pognan, P. (2005). Langues berbères: à la recherche du système
perdu. In Colloque international « Linguistique amazighe: les nouveaux
horizons ». Tétouan.
[21] Taïfi, M. (připravuje se). Dictionnaire raisonné berbère – français. Parlers du
Maroc.
Appendix
creative
commons
Attribution-ShareAlike 3.0 Unported
CREATIVE COMMONS CORPORATION IS NOT A LAW FIRM AND DOES NOT
PROVIDE LEGAL SERVICES. DISTRIBUTION OF THIS LICENSE DOES NOT CREATE
AN ATTORNEY-CLIENT RELATIONSHIP. CREATIVE COMMONS PROVIDES THIS
INFORMATION ON AN “AS-IS” BASIS. CREATIVE COMMONS MAKES NO
WARRANTIES REGARDING THE INFORMATION PROVIDED, AND DISCLAIMS
LIABILITY FOR DAMAGES RESULTING FROM ITS USE.
License
THE WORK (AS DEFINED BELOW) IS PROVIDED UNDER THE TERMS OF THIS CREATIVE
COMMONS PUBLIC LICENSE (“CCPL” OR “LICENSE”). THE WORK IS PROTECTED BY
COPYRIGHT AND/OR OTHER APPLICABLE LAW. ANY USE OF THE WORK OTHER THAN AS
AUTHORIZED UNDER THIS LICENSE OR COPYRIGHT LAW IS PROHIBITED.
BY EXERCISING ANY RIGHTS TO THE WORK PROVIDED HERE, YOU ACCEPT AND
AGREE TO BE BOUND BY THE TERMS OF THIS LICENSE. TO THE EXTENT THIS LICENSE
MAY BE CONSIDERED TO BE A CONTRACT, THE LICENSOR GRANTS YOU THE RIGHTS
CONTAINED HERE IN CONSIDERATION OF YOUR ACCEPTANCE OF SUCH TERMS AND
CONDITIONS.
1. Definitions
a. “Adaptation” means a work based upon the Work, or upon the Work and other pre-existing
works, such as a translation, adaptation, derivative work, arrangement of music or other
alterations of a literary or artistic work, or phonogram or performance and includes
cinematographic adaptations or any other form in which the Work may be recast, transformed, or
adapted including in any form recognizably derived from the original, except that a work that
constitutes a Collection will not be considered an Adaptation for the purpose of this License. For
the avoidance of doubt, where the Work is a musical work, performance or phonogram, the
synchronization of the Work in timed-relation with a moving image (“synching”) will be
considered an Adaptation for the purpose of this License.
b. “Collection” means a collection of literary or artistic works, such as encyclopedias and
anthologies, or performances, phonograms or broadcasts, or other works or subject matter other
than works listed in Section 1(f) below, which, by reason of the selection and arrangement of
their contents, constitute intellectual creations, in which the Work is included in its entirety in
unmodified form along with one or more other contributions, each constituting separate and
independent works in themselves, which together are assembled into a collective whole. A work
that constitutes a Collection will not be considered an Adaptation (as defined below) for the
purposes of this License.
c. “Creative Commons Compatible License” means a license that is listed at
http://creativecommons.org/compatiblelicenses that has been approved by Creative Commons as
being essentially equivalent to this License, including, at a minimum, because that license: (i)
contains terms that have the same purpose, meaning and effect as the License Elements of this
License; and, (ii) explicitly permits the relicensing of adaptations of works made available under
that license under this License or a Creative Commons jurisdiction license with the same
License Elements as this License.
d. “Distribute” means to make available to the public the original and copies of the Work or
Adaptation, as appropriate, through sale or other transfer of ownership.
e.
“License Elements” means the following high-level license attributes as selected by
Licensor and indicated in the title of this License: Attribution, ShareAlike.
f. “Licensor” means the individual, individuals, entity or entities that offer(s) the Work under
the terms of this License.
g. “Original Author” means, in the case of a literary or artistic work, the individual,
individuals, entity or entities who created the Work or if no individual or entity can be identified,
the publisher; and in addition (i) in the case of a performance the actors, singers, musicians,
dancers, and other persons who act, sing, deliver, declaim, play in, interpret or otherwise
perform literary or artistic works or expressions of folklore; (ii) in the case of a phonogram the
producer being the person or legal entity who first fixes the sounds of a performance or other
sounds; and, (iii) in the case of broadcasts, the organization that transmits the broadcast.
h. “Work” means the literary and/or artistic work offered under the terms of this License
including without limitation any production in the literary, scientific and artistic domain,
whatever may be the mode or form of its expression including digital form, such as a book,
pamphlet and other writing; a lecture, address, sermon or other work of the same nature; a
dramatic or dramatico-musical work; a choreographic work or entertainment in dumb show; a
musical composition with or without words; a cinematographic work to which are assimilated
works expressed by a process analogous to cinematography; a work of drawing, painting,
architecture, sculpture, engraving or lithography; a photographic work to which are assimilated
works expressed by a process analogous to photography; a work of applied art; an illustration,
map, plan, sketch or three-dimensional work relative to geography, topography, architecture or
science; a performance; a broadcast; a phonogram; a compilation of data to the extent it is
protected as a copyrightable work; or a work performed by a variety or circus performer to the
extent it is not otherwise considered a literary or artistic work.
i. “You” means an individual or entity exercising rights under this License who has not
previously violated the terms of this License with respect to the Work, or who has received
express permission from the Licensor to exercise rights under this License despite a previous
violation.
j. “Publicly Perform” means to perform public recitations of the Work and to communicate
to the public those public recitations, by any means or process, including by wire or wireless
means or public digital performances; to make available to the public Works in such a way that
members of the public may access these Works from a place and at a place individually chosen
by them; to perform the Work to the public by any means or process and the communication to
the public of the performances of the Work, including by public digital performance; to
broadcast and rebroadcast the Work by any means including signs, sounds or images.
k. “Reproduce” means to make copies of the Work by any means including without limitation
by sound or visual recordings and the right of fixation and reproducing fixations of the Work,
including storage of a protected performance or phonogram in digital form or other electronic
medium.
2. Fair Dealing Rights. Nothing in this License is intended to reduce, limit, or restrict any uses free from
copyright or rights arising from limitations or exceptions that are provided for in connection with the
copyright protection under copyright law or other applicable laws.
3. License Grant. Subject to the terms and conditions of this License, Licensor hereby grants You a
worldwide, royalty-free, non-exclusive, perpetual (for the duration of the applicable copyright) license to
exercise the rights in the Work as stated below:
a. to Reproduce the Work, to incorporate the Work into one or more Collections, and to
Reproduce the Work as incorporated in the Collections;
b. to create and Reproduce Adaptations provided that any such Adaptation, including any
translation in any medium, takes reasonable steps to clearly label, demarcate or otherwise
identify that changes were made to the original Work. For example, a translation could be
marked “The original work was translated from English to Spanish,” or a modification could
indicate “The original work has been modified.”;
c. to Distribute and Publicly Perform the Work including as incorporated in Collections; and,
d. to Distribute and Publicly Perform Adaptations.
e. For the avoidance of doubt:
i. Non-waivable Compulsory License Schemes. In those jurisdictions in which
the right to collect royalties through any statutory or compulsory licensing scheme
cannot be waived, the Licensor reserves the exclusive right to collect such royalties
for any exercise by You of the rights granted under this License;
ii. Waivable Compulsory License Schemes. In those jurisdictions in which the
right to collect royalties through any statutory or compulsory licensing scheme can be
waived, the Licensor waives the exclusive right to collect such royalties for any
exercise by You of the rights granted under this License; and,
iii. Voluntary License Schemes. The Licensor waives the right to collect royalties,
whether individually or, in the event that the Licensor is a member of a collecting
society that administers voluntary licensing schemes, via that society, from any
exercise by You of the rights granted under this License.
The above rights may be exercised in all media and formats whether now known or hereafter devised.
The above rights include the right to make such modifications as are technically necessary to exercise the
rights in other media and formats. Subject to Section 8(f), all rights not expressly granted by Licensor are
hereby reserved.
4. Restrictions. The license granted in Section 3 above is expressly made subject to and limited by the
following restrictions:
a. You may Distribute or Publicly Perform the Work only under the terms of this License. You
must include a copy of, or the Uniform Resource Identifier (URI) for, this License with every
copy of the Work You Distribute or Publicly Perform. You may not offer or impose any terms on
the Work that restrict the terms of this License or the ability of the recipient of the Work to
exercise the rights granted to that recipient under the terms of the License. You may not
sublicense the Work. You must keep intact all notices that refer to this License and to the
disclaimer of warranties with every copy of the Work You Distribute or Publicly Perform. When
You Distribute or Publicly Perform the Work, You may not impose any effective technological
measures on the Work that restrict the ability of a recipient of the Work from You to exercise the
rights granted to that recipient under the terms of the License. This Section 4(a) applies to the
Work as incorporated in a Collection, but this does not require the Collection apart from the
Work itself to be made subject to the terms of this License. If You create a Collection, upon
notice from any Licensor You must, to the extent practicable, remove from the Collection any
credit as required by Section 4(c), as requested. If You create an Adaptation, upon notice from
any Licensor You must, to the extent practicable, remove from the Adaptation any credit as
required by Section 4(c), as requested.
b. You may Distribute or Publicly Perform an Adaptation only under the terms of: (i) this
License; (ii) a later version of this License with the same License Elements as this License; (iii)
a Creative Commons jurisdiction license (either this or a later license version) that contains the
same License Elements as this License (e.g., Attribution-ShareAlike 3.0 US)); (iv) a Creative
Commons Compatible License. If you license the Adaptation under one of the licenses
mentioned in (iv), you must comply with the terms of that license. If you license the Adaptation
under the terms of any of the licenses mentioned in (i), (ii) or (iii) (the “Applicable License”),
you must comply with the terms of the Applicable License generally and the following
provisions: (I) You must include a copy of, or the URI for, the Applicable License with every
copy of each Adaptation You Distribute or Publicly Perform; (II) You may not offer or impose
any terms on the Adaptation that restrict the terms of the Applicable License or the ability of the
recipient of the Adaptation to exercise the rights granted to that recipient under the terms of the
Applicable License; (III) You must keep intact all notices that refer to the Applicable License
and to the disclaimer of warranties with every copy of the Work as included in the Adaptation
You Distribute or Publicly Perform; (IV) when You Distribute or Publicly Perform the
Adaptation, You may not impose any effective technological measures on the Adaptation that
restrict the ability of a recipient of the Adaptation from You to exercise the rights granted to that
recipient under the terms of the Applicable License. This Section 4(b) applies to the Adaptation
as incorporated in a Collection, but this does not require the Collection apart from the
Adaptation itself to be made subject to the terms of the Applicable License.
c. If You Distribute, or Publicly Perform the Work or any Adaptations or Collections, You
must, unless a request has been made pursuant to Section 4(a), keep intact all copyright notices
for the Work and provide, reasonable to the medium or means You are utilizing: (i) the name of
the Original Author (or pseudonym, if applicable) if supplied, and/or if the Original Author and/
or Licensor designate another party or parties (e.g., a sponsor institute, publishing entity, journal)
for attribution (“Attribution Parties”) in Licensor's copyright notice, terms of service or by other
reasonable means, the name of such party or parties; (ii) the title of the Work if supplied; (iii) to
the extent reasonably practicable, the URI, if any, that Licensor specifies to be associated with
the Work, unless such URI does not refer to the copyright notice or licensing information for the
Work; and (iv) , consistent with Section 3(b), in the case of an Adaptation, a credit identifying
the use of the Work in the Adaptation (e.g., “French translation of the Work by Original Author,”
or “Screenplay based on original Work by Original Author”). The credit required by this Section
4(c) may be implemented in any reasonable manner; provided, however, that in the case of a
Adaptation or Collection, at a minimum such credit will appear, if a credit for all contributing
authors of the Adaptation or Collection appears, then as part of these credits and in a manner at
least as prominent as the credits for the other contributing authors. For the avoidance of doubt,
You may only use the credit required by this Section for the purpose of attribution in the manner
set out above and, by exercising Your rights under this License, You may not implicitly or
explicitly assert or imply any connection with, sponsorship or endorsement by the Original
Author, Licensor and/or Attribution Parties, as appropriate, of You or Your use of the Work,
without the separate, express prior written permission of the Original Author, Licensor and/or
Attribution Parties.
d. Except as otherwise agreed in writing by the Licensor or as may be otherwise permitted by
applicable law, if You Reproduce, Distribute or Publicly Perform the Work either by itself or as
part of any Adaptations or Collections, You must not distort, mutilate, modify or take other
derogatory action in relation to the Work which would be prejudicial to the Original Author's
honor or reputation. Licensor agrees that in those jurisdictions (e.g. Japan), in which any
exercise of the right granted in Section 3(b) of this License (the right to make Adaptations)
would be deemed to be a distortion, mutilation, modification or other derogatory action
prejudicial to the Original Author's honor and reputation, the Licensor will waive or not assert,
as appropriate, this Section, to the fullest extent permitted by the applicable national law, to
enable You to reasonably exercise Your right under Section 3(b) of this License (right to make
Adaptations) but not otherwise.
5. Representations, Warranties and Disclaimer
UNLESS OTHERWISE MUTUALLY AGREED TO BY THE PARTIES IN WRITING, LICENSOR
OFFERS THE WORK AS-IS AND MAKES NO REPRESENTATIONS OR WARRANTIES OF ANY
KIND CONCERNING THE WORK, EXPRESS, IMPLIED, STATUTORY OR OTHERWISE,
INCLUDING, WITHOUT LIMITATION, WARRANTIES OF TITLE, MERCHANTIBILITY, FITNESS
FOR A PARTICULAR PURPOSE, NONINFRINGEMENT, OR THE ABSENCE OF LATENT OR
OTHER DEFECTS, ACCURACY, OR THE PRESENCE OF ABSENCE OF ERRORS, WHETHER OR
NOT DISCOVERABLE. SOME JURISDICTIONS DO NOT ALLOW THE EXCLUSION OF IMPLIED
WARRANTIES, SO SUCH EXCLUSION MAY NOT APPLY TO YOU.
6. Limitation on Liability. EXCEPT TO THE EXTENT REQUIRED BY APPLICABLE LAW, IN NO
EVENT WILL LICENSOR BE LIABLE TO YOU ON ANY LEGAL THEORY FOR ANY SPECIAL,
INCIDENTAL, CONSEQUENTIAL, PUNITIVE OR EXEMPLARY DAMAGES ARISING OUT OF
THIS LICENSE OR THE USE OF THE WORK, EVEN IF LICENSOR HAS BEEN ADVISED OF THE
POSSIBILITY OF SUCH DAMAGES.
7. Termination
a. This License and the rights granted hereunder will terminate automatically upon any breach
by You of the terms of this License. Individuals or entities who have received Adaptations or
Collections from You under this License, however, will not have their licenses terminated
provided such individuals or entities remain in full compliance with those licenses. Sections 1, 2,
5, 6, 7, and 8 will survive any termination of this License.
b. Subject to the above terms and conditions, the license granted here is perpetual (for the
duration of the applicable copyright in the Work). Notwithstanding the above, Licensor reserves
the right to release the Work under different license terms or to stop distributing the Work at any
time; provided, however that any such election will not serve to withdraw this License (or any
other license that has been, or is required to be, granted under the terms of this License), and this
License will continue in full force and effect unless terminated as stated above.
8. Miscellaneous
a. Each time You Distribute or Publicly Perform the Work or a Collection, the Licensor offers
to the recipient a license to the Work on the same terms and conditions as the license granted to
You under this License.
b. Each time You Distribute or Publicly Perform an Adaptation, Licensor offers to the recipient
a license to the original Work on the same terms and conditions as the license granted to You
under this License.
c. If any provision of this License is invalid or unenforceable under applicable law, it shall not
affect the validity or enforceability of the remainder of the terms of this License, and without
further action by the parties to this agreement, such provision shall be reformed to the minimum
extent necessary to make such provision valid and enforceable.
d. No term or provision of this License shall be deemed waived and no breach consented to
unless such waiver or consent shall be in writing and signed by the party to be charged with such
waiver or consent.
e. This License constitutes the entire agreement between the parties with respect to the Work
licensed here. There are no understandings, agreements or representations with respect to the
Work not specified here. Licensor shall not be bound by any additional provisions that may
appear in any communication from You. This License may not be modified without the mutual
written agreement of the Licensor and You.
f. The rights granted under, and the subject matter referenced, in this License were drafted
utilizing the terminology of the Berne Convention for the Protection of Literary and Artistic
Works (as amended on September 28, 1979), the Rome Convention of 1961, the WIPO
Copyright Treaty of 1996, the WIPO Performances and Phonograms Treaty of 1996 and the
Universal Copyright Convention (as revised on July 24, 1971). These rights and subject matter
take effect in the relevant jurisdiction in which the License terms are sought to be enforced
according to the corresponding provisions of the implementation of those treaty provisions in the
applicable national law. If the standard suite of rights granted under applicable copyright law
includes additional rights not granted under this License, such additional rights are deemed to be
included in the License; this License is not intended to restrict the license of any rights under
applicable law.
Creative Commons is not a party to this License, and makes no warranty whatsoever in
connection with the Work. Creative Commons will not be liable to You or any party on
any legal theory for any damages whatsoever, including without limitation any general,
special, incidental or consequential damages arising in connection to this license. Notwithstanding the foregoing two (2) sentences, if Creative Commons has expressly identified
itself as the Licensor hereunder, it shall have all rights and obligations of Licensor.
Except for the limited purpose of indicating to the public that the Work is licensed under
the CCPL, Creative Commons does not authorize the use by either party of the trademark
“Creative Commons” or any related trademark or logo of Creative Commons without the
prior written consent of Creative Commons. Any permitted use will be in compliance with
Creative Commons' then-current trademark usage guidelines, as may be published on its
website or otherwise made available upon request from time to time. For the avoidance of
doubt, this trademark restriction does not form part of the License.
Creative Commons may be contacted at http://creativecommons.org/.
NLP, Corpus Linguistics,
Corpus Based Grammar Research
Editors Jana Levická Radovan Garabík
Cover Design by Vladimír Benko
Typeset by Marek Ivančík
Printed by Tribun EU s.r.o.
Gorkého 41, 602 00 Brno, Czech Republic
http://www.librix.eu
ISBN 978-80-7399-875-2
First published by Tribun EU
Brno 2009